Post on 03-Jul-2020
transcript
Algebraic analysis of V-cycle multigrid
and aggregation-based two-grid methods
Artem NAPOV
2010 Bruxelles
v v
v v
v v v
v v v
v v v
v v v v v
v v v v v
v v v v v
v v v v v
v v v v v
Algebraic analysis of V-cycle multigrid
and aggregation-based two-grid methods
Artem Napov
Directeur de these :
Prof. Yvan NOTAY
Membres du joury :
Prof. Robert BEAUWENS
Prof. Anne DELANDTSHEER
Prof. Pierre-Etienne LABEAU
Prof. Yvan NOTAY
Prof. Cornelis W. OOSTERLEE
Prof. Daniel TUYTTENS
Prof. Stefan VANDEWALLE
These presentee
en vue de l’obtention du grade
de Docteur en Sciences de l’Ingenieur
Universite Libre de Bruxelles
Faculte des Sciences Appilquees
Service de Metrologie Nucleaire
Bruxelles
janvier 2010
“Understanding is, after all, what science is all about – and science is a great deal
more than mindless computation.”
Sir Roger Penrose
“Of course everything in computerology is new; that is at once its attraction, and its
weakness.”
James H. Wilkinson
Remerciements / Acknowledgements
Je remercie tout d’abord mon directeur de these Yvan Notay pour sa disponibilite, le
temps qu’il m’a consacre, pour les discussions enrichissantes et pour son encouragement,
mais par dessus tout pour m’avoir fait decouvrir ce vaste et surprenant chantier des idees
en perfectionnement perpetuel qui est l’analyse des methodes numeriques.
Une pensee particuliere va a tous les membres du service de Metrologie Nucleaire qui
ont contribue, chacun a leur maniere, a une ambiance de travail unique dans laquelle j’ai
ete immerge des les premiers jours. Je remercie en particulier mon collegue Nicolas Pauly
pour les discussions enrichissantes sur la science, l’enseignement et sur bien d’autres
choses encore.
Je remercie egalement les membres du laboratoire Ampere de l’ecole centrale de Lyon,
et en particulier Ronan Perrussel, qui m’ont accueilli chaleureusement en leur sein durant
les quelques semaines de sejour a Lyon. Ce sejour fut enrichissant et fructueux et j’ai
l’espoir que notre collaboration se poursuivra dans le futur avec la meme intensite.
I am very grateful to the members of the jury (especially to Cornelius Oosterlee,
Stefan Vandewalle and Robert Beauwens) who have found time to read (and proofread)
the thesis and made many valuable comments and suggestions.
The last but not least. Un grand merci du fond de mon coeur va a toute ma famille:
ma soeur Krystyna, ma mere Iryna et mon pere Oleh. Sans leur aide et leur soutien je
ne serai pas arrive aussi loin.
v
Contents
1 Introduction 11.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Why linear systems? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.3 Linear system solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.4 Multigrid methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41.5 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61.6 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2 Comparison of bounds for V-cycle multigrid 112.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112.2 General setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132.3 Bounds on the V-cycle multigrid convergence factor . . . . . . . . . . . . 15
2.3.1 SSC theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152.3.2 Hackbusch bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . 222.3.3 McCormick’s bound . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.4 Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232.5 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3 When does two-grid optimality carry over to the V-cycle? 373.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373.2 General setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393.3 Theoretical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.3.1 McCormick’s bound . . . . . . . . . . . . . . . . . . . . . . . . . . 413.3.2 Relationship to the two-grid convergence rate . . . . . . . . . . . . 413.3.3 Fourier analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443.3.4 Finite element setting . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 463.4.1 Standard multigrid with 2D Poisson . . . . . . . . . . . . . . . . . 463.4.2 Aggregation-based multigrid for 1D Poisson . . . . . . . . . . . . . 503.4.3 Positive off-diagonal entries . . . . . . . . . . . . . . . . . . . . . . 52
3.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4 Smoothing factor and actual multigrid convergence 55
vii
viii Contents
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 554.2 General setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 574.3 V–cycle analysis and McCormick’s bound . . . . . . . . . . . . . . . . . . 614.4 Rigorous Fourier analysis for SPD problems . . . . . . . . . . . . . . . . . 624.5 Semi-positive definite problems and local Fourier analysis . . . . . . . . . 724.6 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.6.1 Usual prolongations in 2D . . . . . . . . . . . . . . . . . . . . . . . 764.6.2 2D Poisson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5 Algebraic analysis of aggregation-based multigrid 835.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 835.2 Aggregation-based two-grid schemes . . . . . . . . . . . . . . . . . . . . . 855.3 Algebraic analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 875.4 Discrete PDEs with constant and smoothly varying coefficients . . . . . . 94
5.4.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 945.4.2 Constant coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . 965.4.3 Smoothly varying coefficients . . . . . . . . . . . . . . . . . . . . . 965.4.4 Numerical example . . . . . . . . . . . . . . . . . . . . . . . . . . . 975.4.5 Sharpness of the estimate . . . . . . . . . . . . . . . . . . . . . . . 97
5.5 Discrete PDEs with discontinuous coefficients . . . . . . . . . . . . . . . . 1025.5.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1025.5.2 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1065.5.3 Numerical example . . . . . . . . . . . . . . . . . . . . . . . . . . . 1075.5.4 Sharpness of the estimate . . . . . . . . . . . . . . . . . . . . . . . 109
5.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6 Fourier Analysis of aggregation-based two-grid method for edge ele-ment 1136.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1136.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.2.1 Discretized problem . . . . . . . . . . . . . . . . . . . . . . . . . . 1156.2.2 Reitzinger and Schoberl (RS) multigrid . . . . . . . . . . . . . . . 117
6.3 Fourier analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1206.3.1 Model problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1206.3.2 Fourier analysis setting . . . . . . . . . . . . . . . . . . . . . . . . 123
6.4 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1326.4.1 Two-grid method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1326.4.2 Multigrid implementation . . . . . . . . . . . . . . . . . . . . . . . 135
6.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
List of Figures 137
List of Tables 139
Bibliography 141
Chapter 1Introduction
1.1 Preliminaries
In this thesis we consider multigrid methods for the solution of linear systems of equa-
tions. This introductory chapter aims at situating the research material of the thesis
in the general context of numerical analysis and scientific computing. In particular, the
following section sheds some light on (several of the numerous) applications in which
linear systems can arise. A brief overview of solutions techniques for linear systems is
given in Section 1.3. Basic multigrid concepts are introduced in Section 1.4. In Sec-
tion 1.5 we briefly describe the content of the following five chapters, ending up with
some comments on notation.
The reader familiar with basic multigrid concepts can start directly with Section 1.5.
1.2 Why linear systems?
An important number of problems in science and engineering can be formulated in terms
of linear partial differential equations (PDEs). Such equations frequently arise in:
• electrical engineering ,
• computational fluid dynamics (Stokes and Oseen equations) ,
• structural mechanics ,
• transport phenomena ,
• acoustics ,
• chemistry .
To solve numerically these PDE problems, one first performs their discretization; that
is, the initial continuous problem, formulated at every point of the underlying domain,
1
2 Introduction
is reduced to a limited number of equations with usually the same number of unknowns.
If the initial PDE is linear, so are the resulting equations; otherwise, it is a common
practice to linearize the obtained equations using some suitable Newton-like scheme.
In other words, discrete PDEs usually lead to a linear system, stated in vector-matrix
notation as
Ax = b . (1.1)
The main discretization techniques are:
• finite element methods, which use a linear combination of appropriately chosen
shape functions to approximate the solution; the unknowns are the weights of shape
functions and linear system results from application of a minimization principle to
the discretization error [12,76];
• finite volume methods, based on the subdivision of the underlying domain into cells,
on which unknown function(s) (often describing physical quantities) are assumed
constant; linear system is then formed by balance equations that account on sources
inside cells and on the transport of physical quantities between them;
• finite difference methods, which consider unknown function(s) in a given number
of nodes inside or on the boundary of the domain; the linear system arises from
PDE(s) when derivatives of each unknown function are approximated by its dif-
ferences [56,40].
Systems arising from a discretization of PDEs are often sparse; that is, each of their
equations relate together only a small number of unknowns, and the major part of the
entries of A equals zero. It then makes sense to keep in memory only the nonzero
entries and their position in the matrix, which further enables to tackle problems with
an important number of unknowns (107 for a usual PC).
Besides PDE applications, a number of problems are already discrete and formulated
as a linear system of equations. Such problems arise, for instance, in image restoration
or signal processing [43].
1.3 Linear system solvers
The solution of linear system(s) is the most time-consuming process in the majority of
scientific computing applications and therefore should not be neglected. When regular
systems are considered, it can be performed either by direct or by iterative methods.
Direct methods are usually variants of Gaussian elimination. In practice, this latter
is often performed by factorizing the system matrix into a product of lower and upper
triangular matrices (LU factorization), the process being finished by the consecutive
solution of two related triangular systems. Even if the initial system is sparse, the
Introduction 3
triangular factors rarely have the same sparsity: direct methods often have important
memory requirements.
The idea behind iterative methods is to solve the linear system (1.1) approximately
using a suitable procedure, which we formally denote
x = B(b) .
The system is then solved (exactly) if we recover the correction vector e such that
A(x + e) = b ,
or, equivalently,
Ae = r , (1.2)
where r = b−Ax is called residual. This latter equation (also called correction equation)
is equivalent to the initial system (1.1) and can again by solved approximately. The
procedure is repeated until the required precision is reached.
Note that iterative methods rarely give the exact solution of the linear system (1.1).
However, if properly designed, they allow to come closer to the solution at each iteration
step. This feature is particulary relevant since the solution with only a limited accuracy
is often required.
An important characterization of iterative solvers is their optimality with respect to
a given class A(n) of linear system matrices, where n denote the system size. An optimal
iterative method, when applied to systems with system matrix A(n), should have
• its cost per iteration proportional to the system size n ,
• its convergence rate (gain in precision per iteration step) bounded above by a
constant that does not depend on n .
Clearly, if the solution of the linear system (1.1) is determined up to a desired
precision ε with an optimal iterative method, the computational cost is proportional to
n log(ε). Using direct methods for the same purposes amounts to O(n3)
operations if
the matrix is dense (not sparse) and to O(n2)
operations if it arises from discretization
of typical 2-dimensional PDEs [62, p.9] [61, p.14]. Therefore, for system size n large
enough, optimal (and even some suboptimal) iterative methods become more attractive
than direct solvers.
Among the most popular iterative techniques, we should mention:
• Krylov subspace methods, that can be viewed as simple iterative methods where
the approximation B(r) of correction is weighted after each iteration in order to
satisfy some minimization principle. The approximate solution procedure B(·) can
still be chosen freely and is then called preconditioner.
4 Introduction
• Multigrid (multilevel) methods, which we introduce below, have been the first nu-
merical techniques to reach the optimal convergence for usual applications. They
are considered as the most efficient methods for the solution of system arising from
discretization of elliptic PDEs and among the most efficient approaches for other
PDE applications.
• Domain decomposition methods, that correspond to a class of approaches spe-
cially designed for parallel computer architecture. Their main idea is to split the
unknowns into a number of sets such that communication between such sets is
reduced during the solution process.
• Incomplete factorizations (ILU), often used as preconditioners by default for Krylov
subspace methods. The main idea is to reduce the cost and memory requirements
of direct methods that perform complete LU factorization by dropping some en-
tries in the triangular factors. Due to their purely algebraic nature, ILU techniques
can be of interest when applied to problems for which the other methods fail.
For further details on linear system solvers, we refer to corresponding chapters in [24].
Introductory material on iterative methods (including the main variants listed here) can
be found in [54], whereas more advance subjects are treated in [3]. For further informa-
tion on the preconditioning techniques we refer to [7], whereas a broad presentation of
Krylov methods from the historical perspective can be found in [55]
1.4 Multigrid methods
The efficiency of multigrid methods depends on the interplay between its two main
components: smoother and coarse grid correction. The smoother is often a simple
iterative method, and, if used alone, has poor convergence properties. For Poisson-like
problems∂
∂x
(αx∂u
∂x
)+
∂
∂y
(αy∂u
∂y
)+ βu = f (1.3)
the two well known examples are Jacobi and Gauss-Seidel smoothers [61, Chapters
1-2]; both correspond to a linear approximation procedure B(v) = Bv, where B is,
respectively, the diagonal and (up to some permutations) the lower triangular parts of
A. When applied to the linear system (1.1), such schemes reduce the magnitude of
oscillatory modes in the correction e, while keeping the smooth components unchanged.
After several smoothing iterations, the correction becomes geometrically smooth; that is,
it varies slowly from one point to another (see Figure 1.1 for illustration). Other examples
are block smoothers [61, Section 5.1] for anisotropic problems, ILU smoothers [71, 70]
in computational fluid dynamics applications and hybrid smoothers for problems in
electromagnetics [29](see also Chapter 6).
Introduction 5
00.2
0.40.6
0.81
0
0.5
10.1
0.2
0.3
0.4
0.5
0.6
00.2
0.40.6
0.81
0
0.5
10
0.2
0.4
0.6
0.8
00.2
0.40.6
0.81
0
0.5
10
0.2
0.4
0.6
0.8
(a) (b) (c)Figure 1.1: An example of correction e smoothed by Gauss-Seidel scheme; (a) initialcorrection (b) correction after 1 iteration (c) correction after 2 iterations. The corre-sponding linear system A was obtained by discretization of constant-coefficient isotropic
Poisson PDE (1.3) with Dirichlet boundary conditions on rectangular grid 33× 33.
The smooth character of the correction e can then be exploited, approximating it by
a smaller coarse vector ec of size nc < n which still reproduces the essential part of the
correction’s behaviour. This coarse correction vector is obtained by solving a smaller
coarse nc × nc system
Acec = rc (1.4)
that approximates the initial fine correction system (1.2). This solution corresponds to
the second main multigrid ingredient, known as coarse grid correction.
If the coarse grid correction step is performed by a direct solver, its combination with
a smoothing scheme is called two-grid method. Whereas it is often cheaper than a direct
method, the system to be solved is smaller than the fine one only by a modest factor
(4 in usual applications from two-dimensional PDE problems); the two-grid scheme is
therefore still not optimal. The coarse system (1.4) can however be solved approximately
by (recursively) applying γ iterations of the two-grid method; the recursion argument
can be repeated, forming coarser and coarser systems, until a small enough system size
is reached. If γ = 1, the resulting algorithm is called V–cycle whereas if γ = 2, we
talk about W–cycle (these denominations come from the schematic representation of
the recursion calls). Note that if one solves the coarse system (1.4) by γ iterations of a
relevant Krylov scheme using the two-grid method as a preconditioner, one obtains the
so-called K-cycle [49].
So far, we have not specified how to construct the (hierarchy of) coarse system(s)
(1.4). In case of discretized PDE applications, system matrices of various size can often
be generated for the problem at hand. Combining this with geometrical interpolation to
pass from the coarser correction ec to its finer approximation, we obtain the required in-
gredients. This approach is known as geometric multigrid. It is also possible to construct
the multigrid hierarchy in a black box fashion, based only on the knowledge of the system
matrix A. Such setup phase is usually called coarsening and the black-box multigrid
which uses it is algebraic multigrid. Whereas it is slower than its geometric counterpart
6 Introduction
(because of the additional cost of coarsening), algebraic multigrid can be applied to a
variety of problems, even those which have no PDE or geometric background.
An excellent introduction to the multigrid techniques can be found in [19]. For more
details on practical aspects we refer to [61], whereas a more formal presentation can be
found in [67,27].
1.5 Overview
The remaining five chapters of this thesis treat two essentially different subjects: V-
cycle schemes are considered in Chapters 2-4, whereas the aggregation-based coarsening
is analyzed in Chapters 5-6. As a matter of paradox, these two multigrid ingredients,
when combined together, can hardly lead to an optimal algorithm. Indeed, a V-cycle
needs more accurate prolongations than the simple piecewise-constant one, associated
to aggregation-based coarsening. On the other hand, aggregation-based approaches use
almost exclusively piecewise constant prolongations, and therefore need more involved
cycling strategies, K-cycle [49] being an attractive alternative in this respect.
Chapter 2 considers more precisely the well-known V-cycle convergence theories: the
approximation property based analyses by Hackbusch [27] and by McCormick [38] and
the successive subspace correction theory, as presented in [73] by Xu and in [75] by
Yserentant. Under the constraint that the resulting upper bound on the convergence
rate must be expressed with respect to parameters involving two successive levels at a
time, these theories are compared. Unlike [75], where the comparison is performed on
the basis of underlying assumptions in a particular PDE context, we compare directly
the upper bounds. We show that these analyses are equivalent from the qualitative
point of view. From the quantitative point of view, we show that the bound due to
McCormick is always the best one.
When the upper bound on the V-cycle convergence factor involves only two successive
levels at a time, it can further be compared with the two-level convergence factor. Such
comparison is performed in Chapter 3, showing that a nice two-grid convergence (at
every level) leads to an optimal McCormick’s bound (the best bound from the previous
chapter) if and only if a norm of a given projector is bounded on every level.
In Chapter 4 we consider the Fourier analysis setting for scalar PDEs and extend the
comparison between two-grid and V-cycle multigrid methods to the smoothing factor.
In particular, a two-sided bound involving the smoothing factor is obtained that defines
an interval containing both the two-grid and V-cycle convergence rates. This interval
is narrow when an additional parameter α is small enough, this latter being a simple
function of Fourier components.
Chapter 5 provides a theoretical framework for coarsening by aggregation. An upper
bound is presented that relates the two-grid convergence factor with local quantities,
Introduction 7
each being related to a particular aggregate. The bound is shown to be asymptotically
sharp for a large class of elliptic boundary value problems, including problems with
anisotropic and discontinuous coefficients.
In Chapter 6 we consider problems resulting from the discretization with edge finite
elements of 3D curl-curl equation. The variables in such discretization are associated
with edges. We investigate the performance of the Reitzinger and Schoberl algorithm
[52], which uses aggregation techniques to construct the edge prolongation matrix. More
precisely, we perform a Fourier analysis of the method in two-grid setting, showing its
optimality. The analysis is supplemented with some numerical investigations.
All chapters are independent from each other and can be read in any order. We
recommend however the reading of Chapters 2-4 in the ascending order since the results
demonstrated in the earlier chapters are used in the following ones.
Chapters 2 through 5 have appeared as separate papers or reports. Their presenta-
tion have been only slightly modified in this thesis. In particular, Chapter 2 corresponds
to
A. Napov and Y. Notay Comparison of bounds for V-cycle multigrid,
published online in Appl. Numer. Math.
DOI: 10.1016/j.apnum.2009.11.003, 2009,
Chapter 3 is taken from
A. Napov and Y. Notay When does two-grid optimality carry over to the V-
cycle?, accepted for publication in Numer. Lin. Alg. Appl., 2009,
whereas Chapter 4 is a slightly modified version of
A. Napov and Y. Notay Smoothing factor and actual multigrid convergence,
Report GANMN 09-03, Universite Libre de Bruxelles, Brussels, Belgium, 2009,
and Chapter 5 reproduces the content of
A. Napov and Y. Notay Algebraic analysis of aggregation-based multigrid,
Report GANMN 09-04, Universite Libre de Bruxelles, Brussels, Belgium, 2009.
Regarding the Chapter 6, its content is a result of author’s collaboration with Ronan
Perrussel from Laboratoire Ampere, Ecole Centrale de Lyon. The corresponding paper
is still in preparation and the content of this chapter it the author’s contribution to the
common research. Numerical experiments in the multilevel setting are limited here to
the model problem setting; the algebraic multilevel implementation of the presented ap-
proach and the related numerical experiments correspond to the contribution of Ronan
Perrussel and will appear in the final manuscript.
8 Introduction
1.6 Notation
We use bold lowercase Roman letters (e.g., v) to denote vectors and uppercase Roman
(e.g., A) to denote matrices. Capital calligraphic letters (e.g., V) represent vector sub-
spaces, except O, which stands for Landau big Oh symbol, and symbols in Chapter 6,
which are used to denote Fourier block matrices and index sets.
We use I to denote the identity matrix and O the zero matrix. When the dimensions
are not obvious from the context, we write more specifically Im for the m×m identity
matrix, and Om×l for the m× l zero matrix.
For any real α, bαc is the largest integer not greater than α. For any set Γ, |Γ| is
its size. For any real matrix B, R(B) is the range of B and N (B) is its null space; BT
stands for its transpose and BH for its transpose complex conjugate. For any square
real matrix C, ρ(C) is its spectral radius (that is, its largest eigenvalue in modulus),
‖C‖ =√ρ(CTC) is the usual 2–norm and ‖C‖F =
√∑i,j C
2ij the Frobenius norm. For
an SPD matrix D, ‖v‖D =(vTDv
)1/2 = ‖D1/2v‖ is the associated D-norm of a vector
v (if D = A, it is also called energy norm) and
‖C‖D = maxv
‖Cv‖D‖v‖D
= ‖D1/2CD−1/2‖
is the induced matrix D-norm.
We finish this section by giving the list of acronyms and the list of symbols below.
List of Acronyms
Acronym Meaning
ARPACK Arnoldi package [36]
FCG flexible conjugated gradient
GS Gauss-Seidel (smoother)
PDE partial differential equation
RS Reitzinger and Schoberl multigrid method [52]
SPD symmetric positive definite
SSC successive subspace correction
Introduction 9
List of Symbols
Symbol Meaning Reference
A generic system matrix e.g., (1.1)
Ac, Ak coarse grid matrix, coarse grid matrix on level k p.86, p.13
cA approximation property constant in Hackbush’s theory (2.31)
E(k)MG V-cycle multigrid iteration matrix on kth grid e.g., (2.2)
ETG, E(k)TG two-grid iteration matrix (between kth and (k − 1)th grid) e.g., (3.3)
Gk auxiliary matrix inducing a decomposition in SSC theory p.15
h mesh size on a regular grid
J index of the finest level in multilevel setting p.13
K parameter in SSC convergence theory (2.10)
M Chap. 6: mass matrix in edge-element discretization of (6.2) p.116
M (·) Chap. 2-4: equivalent pre– or post–smoothing matrix e.g., (2.3)
n, nk size of A, size of Akn(k) Chap. 5: size of kth aggregate p.86
N Chap. 2-3 and 5-6: number of grid unknowns in one direction
N (·) Chap. 4: I −N (ν)k Ak = (I −R−1
k Ak)ν (4.3)
P, Pk prolongation matrix (from kth and (k − 1)th grid ) e.g., p.13
R smoother matrix of elementary smoothers (e.g., Gauss-Seidel) e.g., p.13
S Chap. 5-6: smoothing iteration matrix e.g., p.119
X Chap. 5-6: equivalent pre– and post–smoothing matrix (5.6), (6.13)
α, β Chap. 4: V-cycle convergence parameters e.g., (4.14),
(4.16)
Chap. 5-6: PDE coefficients (5.37),(6.1)
Γ auxiliary matrix in SSC convergence theory (2.11)
Γk, Γk aggregate k or k p.117, p.125
δ approximation property constant in McCormick’s theory e.g., (2.34)
θ Chap. 2-4, 6: “frequency” in Fourier analysis
µ, µ(k) Chap. 4: smoothing factor (on kth grid) p.60
Chap. 5: two-grid quality (of kth aggregate) p.87, (5.19)
ν number of smoothing steps p.13
πC projector, generally of the form P (P TCP )−1P TC e.g., (3.5)
ω(·) parameter in V-cycle convergence theories (2.4)
ω, ωJac smoother weighting
Ω PDE domain
Chapter 2Comparison of bounds for V-cycle multigrid
Summary
We consider multigrid methods with V-cycle for symmetric positive definite linear sys-
tems. We compare bounds on the convergence factor that are characterized by a constant
which is the maximum over all levels of an expression involving only two consecutive
levels. More particularly, we consider the classical bound by Hackbusch, a bound by
McCormick, and a bound obtained by applying the successive subspace correction con-
vergence theory with so-called a-orthogonal decomposition. We show that the constants
in these bounds are closely related, and hence that these analyses are equivalent from
the qualitative point of view. From the quantitative point of view, we show that the
bound due to McCormick is always the best one. We also show on an example that it
can give satisfactory sharp prediction of actual multigrid convergence.
2.1 Introduction
We consider multigrid methods for solving symmetric positive definite (SPD) n×n linear
systems:
Ax = b. (2.1)
Multigrid methods are based on the recursive use of a two–grid scheme. A basic two–
grid method combines the action of a smoother, often a simple iterative method such
as Gauss-Seidel, and a coarse grid correction, which involves solving a smaller problem
on a coarser grid. A V–cycle multigrid method is obtained when this coarse problem is
solved approximately with 1 iteration of the two–grid scheme on that level, and so on,
until the coarsest level, where an exact solve is performed. Other cycles may be defined,
including the W–cycle based on two recursive applications of the two-grid scheme at
each level; see, e.g., [61].
11
12 Comparison of bounds for V-cycle multigrid
When the system (2.1) stems from the discretization of an elliptic PDE, the V-cycle
multigrid has often optimal convergence properties; that is, the convergence is indepen-
dent of the number of levels and of the mesh discretization parameter h. There are two
classical ways for proving this. One way consists in checking the so-called smoothing and
approximation properties [10,13,26,27,37,38,53]. Another possibility consists in defining
an appropriate subspace decomposition and then analyze the constants involved in the
successive subspace correction (SSC) convergence theory [50, 51, 25, 73, 75, 74]. So far,
these approaches have only been compared (e.g., in [75]) on the basis of the regularity
assumptions that an elliptic boundary value problem should fulfill in order to guarantee
optimal bounds for the multigrid method applied to its finite element discretization.
This allows only qualitative conclusions which are further restricted to a specific con-
text. For instance, such comparison does not cover V-cycle multigrid for structured
linear systems [1]. In fact, a detailed comparison of the convergence theories for V-cycle
is difficult because they may be (and have been) formulated diversely. There is some
freedom in choosing the subspace decomposition for the SSC convergence theory and
there is no unique definition of the smoothing and approximation properties.
The smoothing and approximation property ideas form the basis of the early proofs
[10,13,26] of h-independent V-cycle convergence. For the case when A is SPD, the clas-
sical proof is presented in [27, Theorem 7.2.2] by Hackbusch. The convergence estimate
is then characterized by the approximation property constant cA, which is a maximum
over all levels of an expression involving only two consecutive levels.
An alternative approach has been developed by McCormick in [38] (see also [37,53]).
Here again, the convergence estimate depends on a constant δ which is a minimum over
all levels of an expression involving two consecutive levels.
The SSC convergence theory is more recent and also more general, since by tuning
the choice of the space decomposition one can prove some results for elliptic PDEs
without requiring regularity assumptions [14]. The comparison with other approaches
is not easy because this theory is traditionally formulated in an abstract setting. In this
chapter, we first develop an algebraic formulation of the theory, resulting in a bound
which also depends on freely chosen quantities. Next, we justify that this degree of
freedom seemingly disappears if one adds the constraint that one must be able to assess
the main constant in the bound considering only two levels at a time. Note that this
latter constraint is not only mandatory to develop the comparison with the other two
approaches. It is also very sensible in view of a quantitative analysis, where, as we
illustrate on an example, the Fourier analysis setting is used to numerically calculate
the bounds and compare them with the actual convergence factor.
Transferred back into the original SSC setting, the choice for which this two-level
assessment is possible corresponds to the so-called a-orthogonal decomposition, which
is also the decomposition that has been most extensively used when analyzing multigrid
Comparison of bounds for V-cycle multigrid 13
methods for the class of (H2-) regular problems. Then, the bound depends mainly on
a constant K and, in this chapter, we show that the three constants cA, δ and K are in
fact closely related, namely
K = max(1, cA)
and
δ−1 = c(2)A ,
where c(2)A is a Hackbusch approximation property constant for the number of smoothing
steps being doubled. Hence the three approaches are qualitatively equivalent, in the
sense that they simultaneously succeed or fail to prove optimal convergence. From the
quantitative point of view, it further turns out that McCormick’s bound is the best one.
The reminder of this chapter is organized as follows. In Section 2.2, we state the
general setting of this study and gather the needed assumptions. In Section 2.3, we
develop our algebraic variant of the SSC theory and recall the results of Hackbusch and
McCormick. The comparison is performed in Section 2.4, and an example is analyzed
in Section 2.5.
2.2 General setting
We consider a multigrid method with J + 1 levels (J ≥ 1); index J refers to the finest
level (on which the system (2.1) is to be solved), and index 0 to the coarsest level. The
number of unknowns at level k , 0 ≤ k ≤ J , is denoted nk (hence nJ = n).
Our analysis applies to symmetric multigrid schemes based on the Galerkin principle
for the SPD system (2.1); that is, restriction is the transpose of prolongation and the
matrix Ak at level k , k = J − 1, . . . , 0 , is given by Ak = P Tk Ak+1Pk , where Pk is the
prolongation operator from level k to level k+ 1 ; we also assume that the smoother Rkis SPD and that the number of pre–smoothing steps ν (ν > 0) is equal to the number
of post–smoothing steps. The algorithm for V–cycle multigrid is then as follows.
Multigrid with V–cycle at level k: xn+1 = MG(b, Ak,xn, k)
(1) Relax ν times with smoother Rk : xn ← Smooth(xn, Ak, Rk, ν,b)
(2) Compute residual: rk = b−Akxn(3) Restrict residual: rk−1 = P Tk−1rk(4) Coarse grid correction: if k = 1 , e0 = A−1
0 r0
else ek−1 = MG(rk−1, Ak−1, 0, k − 1)(5) Prolongate coarse grid correction: xn ← xn + Pk−1ek−1
(6) Relax ν times with smoother Rk : xn+1 ← Smooth(xn, Ak, Rk, ν,b)
When applying this algorithm, the error satisfies
A−1k b− xn+1 = E
(k)MG
(A−1k b− xn
)
14 Comparison of bounds for V-cycle multigrid
where the iteration matrix E(k)MG is recursively defined from
E(0)MG = 0 and, for k = 1, 2, . . . , J :
E(k)MG = (I −R−1
k Ak)ν(I − Pk−1(I − E(k−1)
MG )A−1k−1P
Tk−1Ak
)(I −R−1
k Ak)ν(2.2)
(see, e.g., [61, p. 48]). Our main objective is the analysis of the spectral radius of E(J)MG ,
which governs convergence on the finest level. Our analysis makes use of the following
general assumptions.
General assumptions
• n = nJ > nJ−1 > ... > n0 ;
• Pk is an nk+1 × nk matrix of rank nk , k = J − 1, . . . , 0 ;
• AJ = A and Ak = P Tk Ak+1Pk , k = J − 1, . . . , 0 ;
• Rk is SPD and such that ρ(I −R−1k Ak) < 1 , k = J, . . . , 1 .
Note also that most of our results do not refer explicitly to the smoother Rk , but are
stated with respect to the matrices M (ν)k defined from
I − M(ν)k
−1Ak = (I −R−1
k Ak)ν . (2.3)
That is, M (ν)k is the smoother that provides in 1 step the same effect as ν steps with
Rk . The results stated with respect to M(ν)k may then be seen as results stated for the
case of 1 pre– and 1 post–smoothing step, which can be extended to the general case
via the relations (2.3).
Most results depend on the following parameter:
ω(ν) = max
(1 , max
1≤k≤Jmax
wk∈Rnk
wTkAkwk
wTk M
(ν)k wk
). (2.4)
From ρ(I −R−1k Ak) < 1, it follows that ω(1) < 2, whereas (2.3) implies
ω(ν) =
1 if ν is even
1 + (ω(1) − 1)ν if ν is odd.(2.5)
Hence one has also ω(ν) < 2 for all ν. Further, if ω(1) = 1, then ω(ν) = 1 for all ν.
We close this subsection by introducing the projector πAk which plays an important
role throughout this chapter:
πAk = Pk−1A−1k−1P
Tk−1Ak . (2.6)
Comparison of bounds for V-cycle multigrid 15
2.3 Bounds on the V-cycle multigrid convergence factor
2.3.1 SSC theory
We consider the SSC convergence analysis as presented in Theorem 4.4 and Lemma 4.6
in [73], and Theorem 5.1 in [75]. Of course, there are more recent versions of this theory,
e.g., in [74] an identity (known as XZ-identity) is obtained which provides the exact
convergence factor. However, we do not see how to transform these further versions so
that, according to the focus of this chapter, they deliver a bound that could be assessed
considering only two levels at a time (while being significantly different from the bound
given by Theorem 2.1 together with Theorem 2.3). In particular, it seems clear that
the exact convergence factor is a global quantity whose knowledge necessarily involves
information from all levels. Note that SSC ideas are also treated in an algebraic setting
in [65, Section 5], where both the XZ-identity and approximation property approaches
are presented, without however comparing them.
Now, we first develop in Theorem 2.1 below an algebraic version of Theorem 5.1
in [75]. We give a complete proof since this version slightly improves the original for-
mulation, which uses a matrix Γ with the same entries in the strict upper part, but
non-negative entries in the strict lower part and positive entries on the diagonal.
Observe that in Theorem 2.1 below the freedom left in choosing the pseudo restric-
tions Gk corresponds, in the original formulation, to the freedom associated with the
choice of the space decomposition. More precisely, given a set of Gk, k = 0, . . . , J−1, we
can construct a corresponding space decomposition as defined in [75]. In Appendix A
we show that the converse is also true; that is, with any admissible space decomposition
in the original theory, one may associate a set of pseudo restrictions Gk such that The-
orem 2.1 will yield the same bound as Theorem 5.1 in [75], except for the improvement
associated with the refined definition of Γ.
Theorem 2.1. Let E(J)MG be defined by (2.2) with Pk , k = 0, . . . , J−1 , Ak , k = 0, . . . , J ,
and Rk , k = 1, . . . , J , satisfying the general assumptions stated in Section 2.2. For
k = 1, . . . , J , let M (ν)k be defined by (2.3), and set M (ν)
0 = A0 .
Let Gk , k = 0, . . . , J − 1 , be nk × nk+1 matrices, and, for k = 0, . . . , J , let Pk and
Gk be defined by, respectively,
PJ = I
Pk = Pk+1 Pk , k = J − 1, . . . , 0 ,(2.7)
and
GJ = I
Gk = Gk Gk+1 , k = J − 1, . . . , 0 ,(2.8)
16 Comparison of bounds for V-cycle multigrid
with P−1 = G−1 = O .
There holds
ρ(E(J)MG) ≤ 1− 2− ω(ν)
K(ν)(1 + ‖Γ‖)2, (2.9)
where ω(ν) is defined by (2.4),
K(ν) = maxv∈Rn
∑Jk=0 vT GTk (I − Pk−1Gk−1)T M (ν)
k (I − Pk−1Gk−1)GkvvTAv
, (2.10)
and
Γ =
0 γ01 · · · γ0J
0 · · · γ1J
. . ....
0 γ(J−1)J
0
, (2.11)
with, for k = 0, . . . , J − 1 and l = k + 1, . . . , J ,
γkl = maxwk∈Rnk
maxv∈Rn
vT GTl (I − Pl−1Gl−1)T PTl APkwk
(wTk M
(ν)k wk)1/2(vT GTl (I − Pl−1Gl−1)T M (ν)
l (I − Pl−1Gl−1)Glv)1/2.
(2.12)
Moreover,
‖Γ‖ ≤ ω(ν)√J(J + 1)/2 . (2.13)
Proof. In what follows, we omit the superscript (ν) in M(ν)k . We first gather some
useful definitions:
Qk = (I − Pk−1Gk−1)Gk , k = 0, . . . , J ; (2.14)
Tk = Pk(Mk)−1P Tk A , k = 0, . . . , J ; (2.15)
Fk = (I − Tk)(I − Tk−1) · · · (I − T1)(I − T0) , k = 0, . . . , J . (2.16)
In addition we set F−1 = I .
As shown in [65, Proposition 5.1.1] there holds
E(J)MG = (I − TJ)(I − TJ−1) . . . (I − T1)(I − T0)(I − T1) . . . (I − TJ−1)(I − TJ) .
Further, since A−1(I − Tk)T = (I − Tk)A−1 and (I − T0)2 = I − T0 , one has E(J)MG =
FJA−1F TJ A , showing that
ρ(E(J)MG) = ‖FJ‖2A = max
v∈Rn‖FJv‖2AvTAv
. (2.17)
Comparison of bounds for V-cycle multigrid 17
Using this relation, we first show that (2.9) holds if
vTAv ≤ K (1 + ‖Γ‖)2
(J∑l=0
vTF Tl−1ATlFl−1v
)∀v ∈ Rn . (2.18)
Indeed, since ATk = T Tk A and using (2.4), one has, ∀v ∈ Rn ,
||Fk−1v||2A − ||Fkv||2A = (Fk−1v)TAFk−1v − (Fk−1v)T (I − Tk)TA(I − Tk)Fk−1v
= 2vTF Tk−1ATkFk−1v − (Fk−1v)TT Tk ATk(Fk−1v)
= 2vTF Tk−1ATkFk−1v − (Fk−1v)TAPkM−1k P Tk APkM
−1k P Tk A(Fk−1v)
= 2vTF Tk−1ATkFk−1v − (Fk−1v)TAPkM−1k AkM
−1k P Tk A(Fk−1v)
≥ 2vTF Tk−1ATkFk−1v − ω(ν) (Fk−1v)TAPkM−1k P Tk A(Fk−1v)
= (2− ω(ν)) vTF Tk−1ATkFk−1v .
Summing both sides for k = 0, . . . , J shows that, ∀v ∈ Rn ,
‖v‖2A − ‖FJv‖2A ≥ (2− ω(ν))
(J∑l=0
vTF Tl−1ATlFl−1v
),
and it is straightforward to check that this relation, together with (2.18) and (2.17),
implies (2.9).
We now prove (2.18). Observe that, using (2.14), there holds
J∑l=0
PlQl =J∑l=0
Pl(I−Pl−1Gl−1)Gl =J∑l=0
(PlGl − Pl−1Gl−1
)= PJGJ−P−1G−1 = I .
For any v ∈ Rn , one may then decompose vTAv as the sum of two terms (remembering
that F−1 = I):
vTAv =J∑l=0
vTAPlQlv =J∑l=0
vTF Tl−1APlQlv +J∑l=1
vT (I − F Tl−1)APlQlv . (2.19)
In order to prove (2.18), we bound separately the two terms in the right hand side of
(2.19).
Regarding the first term, one has, applying twice the Cauchy-Schwartz inequality,
J∑l=0
vTF Tl−1APlQlv ≤J∑l=0
(vTQTl MlQlv)1/2(vTF Tl−1APlM−1l P Tl AFl−1v)1/2
≤
(J∑l=0
vTQTl MlQlv
)1/2( J∑l=0
vTF Tl−1ATlFl−1v
)1/2
. (2.20)
18 Comparison of bounds for V-cycle multigrid
To estimate the second term, first observe that
I − Fl−1 = I − (I − Tl−1)Fl−2 = (I − Fl−2) + Tl−1Fl−2 = · · · =l−1∑k=0
TkFk−1 .
Therefore,J∑l=1
vT (I − F Tl−1)APlQlv =J∑l=1
l−1∑k=0
vTF Tk−1TTk APlQlv ,
whereas, for any 0 ≤ k < l ≤ J , using successively (2.15) and (2.12) with wk =
M−1k P Tk AFk−1v ,
vTF Tk−1TTk APlQlv = (vTF Tk−1APkM
−1k )P Tk APlQlv
≤ γkl(vTQTl MlQlv)1/2(vTF Tk−1APkM−1k P Tk AFk−1v)1/2
= γkl(vTQTl MlQlv)1/2(vTF Tk−1ATkFk−1v)1/2 .
Hence, since ‖Γ‖ = maxy‖Γy‖‖y‖ = maxx,y
xTΓy‖x‖ ‖y‖ and using the definition (2.11) of Γ,
there holds
J∑l=1
vT (I − F Tl−1)APlQlv ≤J∑l=1
l−1∑k=0
γkl(vTQTl MlQlv)1/2(vTF Tk−1ATkFk−1v)1/2
≤ ‖Γ‖
(J∑l=0
vTQTl MlQlv
)1/2( J∑k=0
vTF Tk−1ATkFk−1v
)1/2
.
Combining the latter result with (2.20), one gets
vTAv ≤ (1 + ‖Γ‖)
(J∑l=0
vTQTl MlQlv
)1/2 ( J∑l=0
vTF Tl−1ATlFl−1v
)1/2
.
Taking the square of both sides, and using (2.10) (which amounts to∑J
l=0 vTQTl MlQlv ≤K vTAv) straightforwardly leads to (2.18), which completes the proof of (2.9).
It remains to prove (2.13). Note that ‖Γ‖ ≤ ‖Γ‖F =(∑J
l=1
∑l−1k=0 γ
2kl
)1/2. Further,
for any 0 ≤ k < l ≤ J and for any w ∈ Rn and wk ∈ Rnk ,
wTQTl PTl APkwk ≤ (wTQTl P
Tl APlQlw)1/2(wT
k PTk APkwk)1/2
= (wTQTl AlQlw)1/2(wTkAkwk)1/2
≤ ω(ν) (wTQTl MlQlw)1/2 (wTkMkwk)1/2 ,
showing that γkl ≤ ω(ν) . The required result straightforwardly follows.
Now, in this chapter, we focus on bounds that can be estimated considering only two
consecutive levels at a time. The following theorem helps to see when the main constant
Comparison of bounds for V-cycle multigrid 19
K(ν) in Theorem 2.1 can be set in that form.
Theorem 2.2. Let Pk and Gk be defined by (2.7) and (2.8) with Pk , k = 0, . . . , J − 1
and Ak , k = 0, . . . , J , satisfying the general assumptions stated in Section 2.2. Then,
for all v ∈ Rn
vTAv =J∑k=0
vT GTk (I − Pk−1Gk−1)TAk(I − Pk−1Gk−1)Gkv (2.21)
+ 2J∑k=0
vT GTk−1PTk−1Ak (I − Pk−1Gk−1) Gkv
=J∑k=0
vT GTk (I − Pk−1Gk−1)TAk(I + Pk−1Gk−1)Gkv . (2.22)
Moreover, if Pk−1Gk−1 is a projector, then
(I − Pk−1Gk−1)TAk(I + Pk−1Gk−1) (2.23)
is nonnegative definite if and only if
Gk−1 = A−1k−1P
Tk−1Ak. (2.24)
Proof. We begin, noting that vTkAkPk−1Gk−1vk = (vTkAkPk−1Gk−1vk)T =
vTk (Pk−1Gk−1)TAkvk holds for all vk ∈ Rnk . Using this relation with vk = Gkv, equa-
tions (2.21) and (2.22) follow from
J∑k=0
vT GTk (I + Pk−1Gk−1)TAk(I − Pk−1Gk−1)Gkv
=J∑k=0
(vT GTk P
Tk APkGkv − vT GTk−1P
Tk−1APk−1Gk−1v
)= vTAv .
Next, (I − Pk−1Gk−1)TAk(I + Pk−1Gk−1) is nonnegative definite if and only if
vTk (I − Pk−1Gk−1)TAk(I + Pk−1Gk−1)vk ≥ 0 ∀vk ∈ Rnk
which in turn is equivalent to
vTkAkvk ≥ vTk (Pk−1Gk−1)TAkPk−1Gk−1vk ∀vk ∈ Rnk ,
20 Comparison of bounds for V-cycle multigrid
this latter being nothing else but
‖Pk−1Gk−1‖Ak ≤ 1.
Hence, if Pk−1Gk−1 is a projector, it has to be orthogonal, and, hence, symmetric with
respect to the ( · , Ak · ) inner product (see [39, Section 5.13]); that is, Pk−1Gk−1 = BkAk
for some symmetric Bk. This implies Gk−1 = Ck−1PTk−1Ak with Ck−1 symmetric. Since
Pk−1 has full rank, Pk−1Gk−1 is then a projector if and only if Ck−1 = A−1k−1; hence the
required result.
Now, consider the definition (2.10) of K(ν). To obtain an expression that can be
assessed considering only two levels at a time, the only possibility we have found is
to express the denominator vTAv as a sum over all levels similar to the sum in the
numerator, and, assuming each term involved to be non-negative, to bound the ratio of
both these sums∑
k ak/∑
k bk by the maximum of the ratios maxk(ak/bk). The first
result of Theorem 2.2 tells us that such a splitting of vTAv always exists, but the second
result tells us that it is exploitable only with Gk−1 = A−1k−1P
Tk−1Ak, since otherwise there
would be negative terms in the sum of the denominator, at least for certain v.1 Note that
these Gk are such that Pk−1Gk−1 = πAk and correspond to the so-called a-orthogonal
decomposition in the original abstract theory. This choice is further analyzed in the
following theorem, where we prove in particular that one has then Γ = 0. Note that
with the original formulation of [75, Theorem 5.1], one could only prove ‖Γ‖ ≤ ω(ν).
Theorem 2.3. Let the assumptions of Theorem 2.1 hold, and let Gk , k = 0, . . . , J − 1 ,
be defined by (2.24). Then, K(ν) and Γ , defined as in Theorem 2.1, satisfy, respectively
K(ν) = max
(1, max
1≤k≤Jmax
wk∈Rnk
wTk (I − πAk)T M (ν)
k (I − πAk)wk
wTk (I − πAk)TAk(I − πAk)wk
)(2.25)
= max
(1, max
1≤k≤Jmax
wk∈Rnk
wTk (I − πAk)T M (ν)
k (I − πAk)wk
wTkAkwk
)(2.26)
and
Γ = 0 , (2.27)
where πAk is defined by (2.6).
1Theorem 3.2 proves this under the additional assumption that PkGk is a projector, but we did notfound any usable bound based on Gk for which PkGk would not be a projector.
Comparison of bounds for V-cycle multigrid 21
Proof. We first prove (2.27). Note that (2.24) implies Gl = A−1l P Tl A, l = 0, ..., J−1.
Hence, for any 0 ≤ k < l ≤ J and all wk ∈ Rnk , v ∈ Rn ,
wTk P
Tk APl(I − Pl−1Gl−1)Glv = wT
k PTk APlA
−1l P Tl Av −wT
k PTk APl−1A
−1l−1P
Tl−1Av
= wTk P
Tk · · ·P Tl−1
(P Tl APlA
−1l
)P Tl Av
−wTk P
Tk · · ·P Tl−2
(P Tl−1APl−1A
−1l−1
)P Tl−1Av
= wTk P
Tk · · ·P Tl−1P
Tl Av −wT
k PTk · · ·P Tl−2P
Tl−1Av
= wTk P
Tk Av −wT
k PTk Av
= 0 ;
γkl = 0 and therefore Γ = 0 readily follows.
We next prove (2.25) and (2.26). Using (2.22) and Pk−1Gk−1 = πAk together with
(I + πAk)T Ak (I − πAk) = (I − πAk)TAk(I − πAk) in the definition (2.10) of K(ν),
one has
K(ν) = maxv∈Rn
∑Jk=0 vT GTk (I − Pk−1Gk−1)T M (ν)
k (I − Pk−1Gk−1)Gkv∑Jk=0 vT GTk (I − Pk−1Gk−1)TAk(I − Pk−1Gk−1)Gkv
(2.28)
= maxv∈Rn
∑Jk=1 vT GTk (I − πAk)T M (ν)
k (I − πAk)Gkv + vT GT0 A0G0v∑Jk=1 vT GTk (I − πAk)TAk(I − πAk)Gkv + vT GT0 A0G0v
≤ max
(1 , max
1≤k≤Jmax
wk∈Rnk
wTk (I − πAk)T M (ν)
k (I − πAk)wk
wTk (I − πAk)TAk(I − πAk)wk
).
This proves that the right hand side of (2.25) is an upper bound on K(ν) ; the right hand
side of (2.26) is a further upper bound since
maxwk∈Rnk
wTk (I − πAk)TM (ν)
k (I − πAk)wk
wTkAkwk
≥ maxvk∈Rnk
vTk (I − πAk)TM (ν)k (I − πAk)vk
vTk (I − πAk)TAk(I − πAk)vk,
as seen by restricting the maximum in the left hand side to wk = (I − πAk)vk (taking
into account that (I − πAk)2 = (I − πAk)).
To prove that the right hand sides of (2.25), (2.26) are also lower bounds on K(ν) ,
let, for k = 0, . . . , J , Qk = (I − Pk−1Gk−1)Gk . Then rewrite (2.28) as
K(ν) = maxv∈Rn
∑Jk=0 vT QTk M
(ν)k Qkv∑J
k=0 vT QTkAkQkv. (2.29)
Since GkPk = Ink for k = 0, . . . , J − 1 , Lemma 2.1 in Appendix B proves that, for
0 ≤ l, k ≤ J with k 6= l ,
QlPlQl = Ql and QkPlQl = Onk×n .
22 Comparison of bounds for V-cycle multigrid
Restricting the maximum in (2.29) to v = PlQlw for some 0 ≤ l ≤ J yields
K(ν) ≥ maxw∈Rn
wT QTl M(ν)l Qlw
wT QTl AlQlw
= maxw∈Rn
wT GTl (I − Pl−1Gl−1)T M (ν)l (I − Pl−1Gl−1)Glw
wT GTl (I − Pl−1Gl−1)TAl(I − Pl−1Gl−1)Glw
= maxwl∈Rnl
wTl (I − Pl−1Gl−1)T M (ν)
l (I − Pl−1Gl−1)wl
wTl (I − Pl−1Gl−1)TAl(I − Pl−1Gl−1)wl
,
the last equality stemming from the fact that Gl, and hence Gl, has full rank (from
(2.24), (2.8), and because Pk has full rank by virtue of our general assumptions). The
conclusion follows because
wTl (I − Pl−1Gl−1)TAl(I − Pl−1Gl−1)wl = wT
l (I − πAl)TAl(I − πAl)wl
= wTl (Al −AlPl−1A
−1l−1P
Tl−1Al)wl
≤ wTl Alwl .
2.3.2 Hackbusch bound
The bound from [27, Theorem 7.2.2] is recalled in the following theorem. Note that
this analysis requires ω(ν) = 1. This condition is however not too restrictive since the
smoother can be scaled to satisfy it. Note also that, according to (2.5), ω(ν) = 1 always
holds for ν even, and that ω(1) = 1 entails ω(ν) = 1 for all ν.
Theorem 2.4. Let E(J)MG be defined by (2.2) with Pk , k = 0, . . . , J−1 , Ak , k = 0, . . . , J ,
and Rk , k = 1, . . . , J , satisfying the general assumptions stated in Section 2.2. For
k = 1, . . . , J , let M (ν)k and ω(ν) be defined, respectively, by (2.3) and (2.4).
Then, if ω(ν) = 1,
ρ(E(J)MG) ≤
c(ν)A
c(ν)A + 2
, (2.30)
where
c(ν)A = max
1≤k≤Jmax
vk∈Rnk
vTk (A−1k − Pk−1A
−1k−1P
Tk−1)vk
vTk M(ν)k
−1vk
. (2.31)
Moreover, if ω(1) = 1,
ρ(E(J)MG) ≤
c(1)A
c(1)A + 2ν
. (2.32)
Note that Theorem 7.2.2 in [27] considers only (2.32). The bound (2.30) is a straight-
forward extension (through the replacement of M (1)k = Rk by M (ν)
k ) that will make easier
the comparison with other approaches. It is not really useful in practice since, as will
be seen, (2.32) is always better than (2.30). Note, however, that (2.30) is more general
since one may have ω(ν) = 1 while ω(1) > 1.
Comparison of bounds for V-cycle multigrid 23
Note also that in [27] some bounds based on cA are also proved for the W and two-
grid cycle, that are better than those obtained by using just the V-cycle bound as a
worst case estimate.
2.3.3 McCormick’s bound
We recall in the following theorem the bound obtained in [38, Lemma 2.3, Theorem 3.4
and Section 5] (see also [37], or [53] for an alternative proof).
Theorem 2.5. Let E(J)MG be defined by (2.2) with Pk , k = 0, . . . , J−1 , Ak , k = 0, . . . , J ,
and Rk , k = 1, . . . , J , satisfying the general assumptions stated in Section 2.2. For
k = 1, . . . , J , let M (ν)k be defined by (2.3).
Then,
ρ(E(J)MG) ≤ 1− δ(ν) , (2.33)
where
δ(ν) = min1≤k≤J
minvk∈Rnk
‖vk‖2Ak − ‖(I − M(ν)k
−1Ak)vk‖2Ak
‖(I − πAk)vk‖2Ak(2.34)
with πAk defined by (2.6).
Moreover,
δ(ν)−1 ≤ 1ν
(δ(1)−1
+ ν − 1). (2.35)
2.4 Comparison
We first state our main result, which relates the constants K(ν) , c(ν)A and δ(ν) .
Theorem 2.6. Let K(ν) , c(ν)A and δ(ν) be defined respectively by (2.25), (2.31) and
(2.34) where Pk , k = 0, . . . , J − 1 , Ak , k = 0, . . . , J , and Rk , k = 1, . . . , J satisfy
the general assumptions stated in Section 2.2. For k = 1, . . . , J , let M (ν)k be defined by
(2.3).
Then
K(ν) = max( 1, c(ν)A ) , (2.36)
and
δ(ν) =1
c(2ν)A
. (2.37)
Proof. Let
Pk = A1/2k Pk−1A
−1/2k−1 , k = 1, . . . , J .
24 Comparison of bounds for V-cycle multigrid
One has
c(ν)A = max
1≤k≤Jmaxv∈Rnk
vT (A−1k − Pk−1A
−1k−1P
Tk−1)v
vT M(ν)k
−1v
= max1≤k≤J
maxv∈Rnk
vT (I −A1/2k Pk−1A
−1k−1P
Tk−1A
1/2k )v
vTA1/2k M
(ν)k
−1A
1/2k v
= max1≤k≤J
maxv∈Rnk
vT (I − PkP Tk )v
vTA1/2k M
(ν)k
−1A
1/2k v
= max1≤k≤J
maxv∈Rnk
vT M(ν)k
1/2A−1/2k (I − PkP Tk )2A
−1/2k M
(ν)k
1/2v
vTv
= max1≤k≤J
maxv∈Rnk
vT (I − PkP Tk )A−1/2k M
(ν)k A
−1/2k (I − PkP Tk )v
vTv.
Since (I − PkP Tk )A−1/2k = (I − A1/2
k Pk−1A−1k−1P
Tk−1A
1/2k )A−1/2
k = A−1/2k (I − πAk)T , this
leads to
c(ν)A = max
1≤k≤Jmaxv∈Rnk
vT (I − πAk)T M (ν)k (I − πAk)v
vTAkv,
hence (2.36).
On the other hand, observing that M (2ν)k satisfies
I − M(2ν)k
−1Ak = (I − M
(ν)k
−1Ak)2 , k = 1, . . . , J ,
one has
δ(ν) = min1≤k≤J
minv∈Rnk
||v||2Ak − ||I − M(ν)k
−1Akv||2Ak
||(I − πAk)v||2Ak
= min1≤k≤J
minv∈Rnk
vTAkv − vT(I − M
(ν)k
−1Ak
)TAk
(I − M
(ν)k
−1Ak
)v
vT (I − πAk)TAk(I − πAk)v
= min1≤k≤J
minv∈Rnk
vTAkv − vTAk
(I − M
(ν)k
−1Ak
)2
v
vT (I − πAk)TAk(I − πAk)v
= min1≤k≤J
minv∈Rnk
vTAkv − vTAk
(I − M
(2ν)k
−1Ak
)v
vT (I − πAk)TAk(I − πAk)v
= min1≤k≤J
minv∈Rnk
vTAk M(2ν)k
−1Akv
vTAk(I − πAk)v
= min1≤k≤J
minv∈Rnk
vT M (2ν)k
−1v
vT (I − πAk)A−1k v
=1
c(2ν)A
.
Comparison of bounds for V-cycle multigrid 25
We are now ready to compare the bounds (2.9), (2.30), (2.32) and (2.33). This is
done in the following theorem.
Theorem 2.7. Let E(J)MG be defined by (2.2) with Pk , k = 0, . . . , J−1 , Ak , k = 0, . . . , J ,
and Rk , k = 1, . . . , J , satisfying the general assumptions stated in Section 2.2.For
k = 1, . . . , J , let M (ν)k and ω(ν) be defined, respectively, by (2.3) and (2.4). Moreover,
let K(ν) , c(ν)A and δ(ν) be defined respectively by (2.25), (2.31) and (2.34).
Then
ρ(E(J)MG) ≤ 1− δ(ν) ≤ 1− 2− ω(ν)
K(ν). (2.38)
Further, if ω(ν) = 1,
ρ(E(J)MG) ≤ 1− δ(ν) ≤
c(ν)A
c(ν)A + 2
, (2.39)
and, if ω(1) = 1,
ρ(E(J)MG) ≤ 1− δ(ν) ≤
c(1)A
c(1)A + 2ν
≤c
(ν)A
c(ν)A + 2
. (2.40)
Moreover,
1− 2− ω(ν)
K(ν)≤ 1− 2− ω(ν)
2δ(ν) , (2.41)
and, if ω(ν) = 1,c
(ν)A
c(ν)A + 2
≤ 1δ(ν) + 1
= 1− δ(ν)
δ(ν) + 1. (2.42)
Proof. Let us first prove two intermediate results:
c(ν)A
2≤ c(2ν)
A ≤c
(ν)A
2− ω(ν)(2.43)
and, if ω(µ) = 1,c
(µ)A
ν≤ c(µν)
A ≤ 1ν
(c
(µ)A + ν − 1
), µ ∈ N+
0 . (2.44)
The first intermediate result (2.43) follows from
M(2ν)k = M
(ν)k
(2 M (ν)
k −Ak)−1
M(ν)k
combined with
2vTk M(ν)k vk ≥ 2vTk M
(ν)k vk − vTkAkvk ≥ (2− ω(ν))vTk M
(ν)k vk , ∀vk ∈ Rnk .
We prove the second intermediate result (2.44) for µ = 1; its generalization to µ > 1
is performed replacing Rk by M(µ)k in the proof below. First, the right inequality (2.44)
26 Comparison of bounds for V-cycle multigrid
is a consequence of (2.35) since, using (2.37) one has
c(ν)A = δ(ν/2)−1 ≤ 1
ν
(δ(1/2)−1
+ ν − 1)
=1ν
(c
(1)A + ν − 1
)where δ(1/2) corresponds to the V-cycle algorithm with a smoother Rk such that
I −R−1k Ak = (I − R−1
k Ak)2 .
Such Rk is indeed well defined since ω(1) = 1 entails that I −A1/2k R−1
k A1/2k is symmetric
nonnegative definite. On the other hand, the left inequality (2.44) is a straightforward
consequence of
vTk M(ν)k
−1vk ≤ ν vTkR
−1k vk , ∀vk ∈ Rnk
which we prove as follows. This relation holds if and only if
vTkA1/2k M
(ν)k
−1A
1/2k vk ≤ ν vTkA
1/2k R−1
k A1/2k vk , ∀vk ∈ Rnk
which, in view of (2.3) and when ω(1) = 1, is satisfied if
1− (1− x)ν ≤ νx ∀x ∈ [0, 1];
that is, if, ∀λ = 1− x ∈ [0, 1),1− λν
1− λ≤ ν ,
which is readily checked from 1−λν1−λ =
∑ν−1i=0 λ
i < ν .
Now, the second inequality (2.38) follows from the right inequality (2.43) combined
with (2.36) and (2.37). The second inequalities (2.39) and (2.40) are equivalent to,
respectively
c(ν)A c
(2ν)A ≥ (c(ν)
A + 2)(c(2ν)A − 1)
and
c(1)A c
(2ν)A ≥ (c(1)
A + 2ν)(c(2ν)A − 1) .
These inequalities follow from the right inequality (2.44), used with (µ , ν) = (ν , 2) and
(µ , ν) = (1 , 2ν), respectively, combined with (2.37). Next, the last inequality of (2.40)
is a consequence of the left inequality of (2.44) used with (µ , ν) = (1 , ν). Finally,
inequalities (2.41) and (2.42) follow from the left inequality (2.43) combined with (2.37)
and (2.36), because δ(ν)−1 ≥ 1, as may be seen from
δ(ν)−1= c(2ν)
= max1≤k≤J
maxwk∈Rnk
wTk (I − πAk)T M (2ν)
k (I − πAk)wk
wTk (I − πAk)TAk(I − πAk)wk
Comparison of bounds for V-cycle multigrid 27
≥ 1ω(2ν)
max1≤k≤J
maxwk∈Rnk
wTk (I − πAk)T M (2ν)
k (I − πAk)wk
wTk (I − πAk)T M (2ν)
k (I − πAk)wk
= 1 .
From (2.38), (2.39) and (2.40), one sees that McCormick’s bound is always the best
one, whereas inequalities (2.41) and (2.42) show that all approaches are nevertheless
qualitatively equivalent, since they give bounds which, at worst, correspond to Mc-
Cormick’s bound with main constant smaller by a modest factor.
2.5 Example
We consider the linear system resulting from the 9-point finite difference discretization
of the two-dimensional Poisson problem
−∆u = f in Ω = (0, 1)× (0, 1)
u = 0 in ∂Ω
on a uniform grid of mesh size h = 1/NJ in both directions. The matrix corresponds
then, up to some scaling factor, to the following nine point stencil−1 −1 −1
−1 8 −1
−1 −1 −1
. (2.45)
We assume NJ = 2JN0 for some integer N0 , allowing J steps of regular geometric
coarsening. We consider prolongations in form of the standard interpolation associated
with bilinear finite element basis functions. The restriction P Tk corresponds then to “full
weighting”, as defined in, e.g. [61] 2. With these choices, the stencil (2.45) is preserved
throughout all grids (up to some unimportant scaling factor), and c(ν)A may be assessed
by analyzing
maxwk
wTk (I − πAk)TM (ν)
k (I − πAk)wk
wkAkwk(2.46)
for a matrix Ak corresponding to stencil (2.45) applied on a grid with mesh size hk =
1/Nk . Considering two successive grids is therefore sufficient, and, to alleviate notation,
we let N = Nk , A = Ak , M (ν) = M(ν)k , P = Pk−1 , Ac = Ak−1 = P TAP and
πA = πAk = PA−1c P TA .
2up to some scaling factor; the scalings of the prolongation and restriction are unimportant whenusing coarse grid matrices of the Galerkin type.
28 Comparison of bounds for V-cycle multigrid
To assess (2.46), we resort to Fourier analysis. The eigenvectors of A are, for m, l =
1, . . . , N − 1, the functions
u(N)m,l = sin(mπx) sin(lπy)
evaluated at the grid points. The eigenvalue corresponding to u(N)m,l is
λ(N)m,l = 4(3sm + 3sl − 4smsl) (2.47)
where
sm = sin2(mπ/2N) , sl = sin2(lπ/2N) . (2.48)
The prolongation P satisfies (see, e.g., [61, p. 87])
P T
u
(N)m,l
u(N)N−m,N−l
−u(N)N−m,l
−u(N)m,N−l
= 4
(1− sm)(1− sl)
smsl
sm(1− sl)(1− sm)sl
u
(N/2)m,l
for 1 ≤ m, l ≤ N/2 − 1 , with P Tu(N)m,l = 0 for m = N/2 or m = N/2 . Expressed in
the Fourier basis (that is, in the basis of eigenvectors of A), I − πA is therefore block
diagonal with, for 1 ≤ m, l ≤ N/2− 1 , 4× 4 blocks
(I − πA)m,l = I4 − Pm,l(A
(c)m,l
)−1P Tm,lAm,l (2.49)
where
P Tm,l = 4(
(1− sm)(1− sl) smsl sm(1− sl) (1− sm)sl)
Am,l = diag(λ
(N)m,l , λ
(N)N−m,N−l , λ
(N)m,N−l , λ
(N)N−m,l
)A
(c)m,l = P Tm,lAm,lPm,l = 64
(3sm(1− sm) + 3sl(1− sl)− 16sl(1− sl)sm(1− sm)
).
For m = N/2 , 1 ≤ l ≤ N/2 − 1 and l = N/2 , 1 ≤ m ≤ N/2 − 1 , (I − πA)m,l = I2 is a
2× 2 identity block, whereas (I − πA)N2, N
2= 1 reduces to the scalar identity. If M (ν) in
the Fourier basis has the same block diagonal structure, we are left with the analysis of
ρm,l = ρ(
(I − πA)Tm,lM(ν)m,l(I − πA)m,lA
−1m,l
). (2.50)
Now, we consider more specifically damped Jacobi smoothing; that is Rk =
ω−1Jacdiag(A) = ω−1
Jac 8 I , with ωJac ∈ ( 0 , 4/3 ) to ensure ω(1) = (3/2)ωJac < 2. Then,
for any number of pre– and post–smoothing steps ν , M (ν) is diagonal in the Fourier
basis, with diagonal entries depending on the eigenvalues of A ; that is (see (2.47)),
Comparison of bounds for V-cycle multigrid 29
depending on sm and sl .To obtain grid independent bounds, it is then interesting to
consider ρm,l = ρ(sm, sl) as a function of sm , sl , and to let these parameters vary
continuously in [0, 1] , excluding the corner points where sm(1−sm) = sl(1−sl) = 0 ,
which correspond to singularities. For all ν, ρ(sm, sl) has the following symmetries:
ρ(sm, sl) = ρ(1−sm, sl) = ρ(sm, 1−sl) = ρ(1−sm, 1−sl) . Further, numerical investi-
gations reveal that the maximum on the considered domain is located at the boundary,
i.e., corresponds to, e.g., sm = 0 . Because of the symmetries it is sufficient to analyze
this latter case. One may check that ρ(0, sl) is the largest eigenvalue in modulus of
14
slµ1+slµ4
3 0 0 − slµ1+slµ4
3
0 µ2
3−(1−sl) 0 0
0 0 µ3
3−sl 0
−µ1(1−sl)+µ4(1−sl)3 0 0 (1−sl)µ1+(1−sl)µ4
3
,
where µii=1,...,4 are the 4 diagonal entries of M (ν)kl , given by
µi =(Am,l)i,i
1− (1− ωJac2 (Am,l)i,i)ν
.
Thus
ρ(0, sl) = max(
µ3
3− sl,
µ2
3− (1− sl),µ1 + µ4
3
),
and, injecting the expressions of µi,
ρ(0, sl) = max(
11− (1− ωJac
2 (3− sl))(ν),
11− (1− ωJac
2 (2 + sl))ν,
sl
1− (1− 3ωJac2 sl)ν
+1− sl
1− (1− 3ωJac2 (1− sl))ν
).
Note that for sl → 0 the third term is larger that the maximum over sl of the first and
the second; hence
ρ(0, sl) ≤ supsl∈(0,1)
(sl
1− (1− 3ωJac2 sl)ν
+1− sl
1− (1− 3ωJac2 (1− sl))ν
). (2.51)
The right hand side of (2.51) is in fact independent of sl for ν = 1, and, for ν = 2
and ν = 4, one may check, using elementary function analysis (see Appendix B), that
the supremum is reached for sl → 0, 1 . Hence
c(ν)A ≤
23νωJac
+1
1− (1− 3ωJac2 )ν
, ν = 1, 2, 4. (2.52)
Using the relation (2.52) as an equality, we can evaluate the different bounds. This is
30 Comparison of bounds for V-cycle multigrid
ωJac ω(1) c(1)A c
(2)A
c(1)A
c(1)A +2
1− 2−ω(1)
K(1) 1− δ(1) ρ(E(J)MG)
1/2 1 2.666 1.733 0.571 0.626 0.423 0.3982/3 1 2 1.5 0.5 0.5 0.333 0.2711 1.5 1.333 1.666 (*) 0.5 0.387 0.251
Table 2.1: Convergence factor of V–cycle (for N0 = 2 and J = 6) and the correspond-ing bounds for ν = 1; (*) the quantity exists, but does not correspond to the bound,
since ω(1) > 1.
ωJac ω(2) c(1)A c
(2)A c
(4)A
c(1)A
c(1)A +4
c(2)A
c(2)A +2
1− 2−ω(2)
K(2) 1− δ(2) ρ(E(J)MG)
1/2 1 2.666 1.733 1.337 0.4 0.4 0.423 0.252 0.1872/3 1 2 1.5 1.25 0.333 0.333 0.333 0.2 0.1211 1 1.333 1.666 1.233 (*) 0.25 0.4 0.189 0.091
Table 2.2: Convergence factor of V–cycle (for N0 = 2 and J = 6) and the correspond-ing bounds for ν = 2; (*) the quantity exists, but does not correspond to the bound,
since ω(1) > 1.
done in Table 2.1 and 2.2 for different number ν of smoothing steps, where we also com-
pare the bounds with the actual convergence factor. One sees that McCormick’s bound
is indeed the best one and, further, that it gives in the considered cases a satisfactory
sharp prediction of actual multigrid convergence.
2.6 Conclusion
We have considered different bounds on the V-cycle multigrid convergence factor, each
depending on a parameter given by the maximum over all levels of a expression defined
on two levels only. More precisely, we have considered the bound in [27, Theorem 7.2.2]
by Hackbusch, the result [38, Lemma 2.3, Theorem 3.4 and Section 5] of McCormick
and the Successive Subspace Correction theory [73, Theorem 4.4 and Lemma 4.6], [75,
Theorem 5.1] used with a-orthogonal decomposition. Regarding the latter approach,
it has been adapted here to the algebraic framework and slightly improved. We have
sown that the main parameters of these three theories are related to each other and
that the corresponding bounds are equivalent from the qualitative point of view; that is,
they simultaneously succeed or fail to prove an optimal convergence for a given problem.
From the quantitative viewpoint, we have proved that the bound of McCormick is the
sharpest, and, further, that it leads to an accurate convergence estimate at least for a
typical example.
Comparison of bounds for V-cycle multigrid 31
Appendix A
We first show that Theorem 5.1 in [75] particularized to the matrix case (that is, applied
to the case of matrix operators in Rn with a(v,w) = (v, Aw) = vTAw) yields the same
bound as Theorem 2.1 (except for the additional refinement in the definition of ‖Γ‖),provided that one has Wk = R(Pk) and Vk = R(PkGk − Pk−1Gk−1), where Pk and Gk
refer to the notation in Theorem 2.1, and Wk,Vk to notation in [75].
Firstly, note that Theorem 5.1 provides a bound on the energy norm of product
iteration matrices of the form (2.16), where
Tk = B+k QkA , (2.53)
B+k being a matrix corresponding to a invertible operator onto Wk , and Qk being the
orthogonal projector on the subspace Wk = R(Pk) ; that is, Qk = Pk(P Tk Pk)−1P Tk .
It then follows that the definition (2.53) matches (2.15) by setting B+k = PkM
−1k P Tk .
Observe also that, ∀wk ∈ Wk,
zk = B+k wk ⇔ wk = Pk(P Tk Pk)
−1Mk(P Tk Pk)−1P Tk zk .
Hence
Bk = Pk(P Tk Pk)−1Mk(P Tk Pk)
−1P Tk (2.54)
is the proper inverse of B+k onto Wk.
Next, the bound on ‖FJ‖2A in [75] is based on the decomposition of any vector v ∈ Rn
as
v =J∑k=0
vk
where vk ∈ Vk. With Vk = R(PkGk − Pk−1Gk−1), it means
vk = Pk(I − Pk−1Gk−1)Gkv = (PkGk − Pk−1Gk−1)v. (2.55)
Then, the bound in [75] is
‖FJ‖2A ≤ 1− 2− ωK1(1 +K2)2
, (2.56)
where K1 is such that
J∑k=0
(Bkvk,vk) ≤ K1vTAv ∀v ∈ Rn , (2.57)
32 Comparison of bounds for V-cycle multigrid
where ω satisfy
(Awk,wk) ≤ ω(Bkwk,wk) ∀wk ∈ Wk , k = 1, ..., J , (2.58)
and where K2 = ‖Γ‖, with Γ = (γkl) being the (J+1)×(J+1) matrix whose coefficients
are such that
(Awk,vl) ≤ γkl(Bkwk,wk)1/2(Blvl,vl)1/2 ∀vk ∈ Vk , wk ∈ Wk (2.59)
for k ≤ l, and γkl = γlk for k > l.
With (2.54) and (2.55), it is easy to recognize that K(ν) in (2.10) is the best constant
K1 satisfying (2.57). On the other hand, “ ∀wk ∈ Wk ” means “ for all wk = Pkw with
w ∈ Rn ” and “ ∀vk ∈ Vk ” means “ for all vk = Pk(I − Pk−1Gk−1)Gkv with v ∈ Rn ”.
Hence, for k < l, γkl in (2.12) is the best γkl satisfying (2.59). Further, using the same
arguments, we see that ω(ν) is the best choice for ω. Therefore, the equivalence between
the bound (2.56) in [75] and (2.9) is proved, except for the additional refinement showing
that the lower triangular part of Γ can be set to zero.
We next show that with any admissible choice of Vk, one may associate valid Gk, k =
0, ..., J such that Vk = R(PkGk − Pk−1Gk−1) (setting P−1 = G−1 = On0×n0). In other
words, any bound from Theorem 5.1 in [75] obtained using a particular decomposition
can also be obtained via (2.9) (up to some additional refinement in the definition of ‖Γ‖)using a particular set of matrices Gk.
We begin the proof letting
Xk = V0 ⊕ V1 ⊕ . . .⊕ Vk .
Observe that the proposition holds if, given X0 ⊂ X1 ⊂ . . . ⊂ XJ = Rn, one can find Gk,
k = 0, ..., J such that
R(PkGk) = Xk (2.60)
and
R(PkGk − Pk−1Gk−1) ∩ R(Pk−1Gk−1) = 0.
The latter equality is checked if, for all v, w ∈ Rn,
(PkGk − Pk−1Gk−1
)v = Pk−1Gk−1w ⇒
(PkGk − Pk−1Gk−1
)v = Pk−1Gk−1w = 0 ;
that is, since Pk has full rank, if
(I − Pk−1Gk−1)(Gkv
)= Pk−1Gk−1
(Gkw
)⇒ (I − Pk−1Gk−1)
(Gkv
)= Pk−1Gk−1
(Gkw
)= 0.
(2.61)
Comparison of bounds for V-cycle multigrid 33
This proposition is true when Pk−1Gk−1 is a projector (note that P−1G−1 = On0×n0 is
a projector as well). The right equalities (2.61) follow then from the multiplication of
(2.61) by (I − Pk−1Gk−1) and Pk−1Gk−1, respectively.
We now assume that Gj has been constructed properly for j = J, ..., k + 1 (which
holds trivially for j = J − 1), and show that one can construct Gk such that
R(PkGkGk+1) = Xk (2.62)
while satisfying the constraint
GkPkGk = Gk , (2.63)
yielding the required result by induction, since (2.63) implies (PkGk)2 = PkGk.
Let mk = dim(Xk). Observe that W0 ⊂ . . . ⊂ Wk implies mk ≤ dim(Wk) = nk.
Hence (2.62) holds if Gk = R(Gk) is a prescribed mk-dimensional subspace of Rnk whose
image by Pk is Xk. Let Hk be an nk ×mk matrix whose columns form a basis of this
subspace. We search for Gk of the form
Gk = HkZk ,
where Zk is an mk × nk+1 matrix of rank mk. Then (2.62) holds if ZkGk+1 has rank
mk, which is ensured if R(Gk+1) contains an mk-dimensional subspace complementary
to N (Zk) (see [39, p. 199]). Note that dim(R(Gk+1)) = dim(Xk+1) ≥ mk, hence there
exists at least one mk-dimensional subspace Gk of R(Gk+1), and we shall enforce the
null space of Zk to be complementary to Gk.Consider now the constraint (2.63). With the given form of Gk, it is satisfied when
ZkPkHk = Imk ;
that is, according to the terminology in [6], if Zk is a 1, 2-inverse of PkHk. As shown
in [6, p. 59], given any subspace Sk complementary to Tk = R(PkHk) there exist such a
1, 2-inverse having Sk as a null space.
Hence the required result is proven if one can always find Sk complementary to both
Gk and Tk. This, in turn, is true since Gk and Tk are subspaces of the same dimension
of a finite dimensional space Rnk , see [34].
34 Comparison of bounds for V-cycle multigrid
Appendix B
Lemma 2.1. Let Pk , k = 0, . . . , J − 1 be nk+1× nk matrices of rank nk with n = nJ >
nJ−1 > · · · > n0 . Let Gk , k = 0, . . . , J − 1 be nk+1 × nk matrices such that
Gk Pk = Ink .
Set P−1 = G−1 = On0×n0 and let, for k = 0, . . . , J , Pk be defined by (2.7), Gk be defined
by (2.8), and Qk = (I − Pk−1Gk−1)Gk .
There holds, for 0 ≤ l, k ≤ J with k 6= l ,
QkPkQk = Qk and QlPkQk = Onl×n .
Proof. Note that Gk Pk = Ink implies GkPk = Ink . The first statement follows then
from
(I − Pk−1Gk−1)GkPk(I − Pk−1Gk−1) = (I − Pk−1Gk−1)(I − Pk−1Gk−1)
= I − Pk−1Gk−1 .
To prove the second statement, we consider two cases. If l > k ,
(I − Pl−1Gl−1)GlPk = (I − Pl−1Gl−1)Gl · · ·GJ−1PJ−1 · · ·PlPl−1 · · ·Pk
= (I − Pl−1Gl−1)Pl−1 · · ·Pk
= Pl−1(I −Gl−1Pl−1)Pl−2 · · ·Pk
= Onl×nk ,
whereas, if l < k ,
GlPk(I − Pk−1Gk−1) = Gl · · ·Gk−1Gk · · ·GJ−1PJ−1 · · ·Pk(I − Pk−1Gk−1)
= Gl · · ·Gk−1(I − Pk−1Gk−1)
= Gl · · ·Gk−2(I −Gk−1Pk−1)Gk−1
= Onl×nk .
Comparison of bounds for V-cycle multigrid 35
Appendix C
In this appendix we outline for even values of ν the proof of the following identity
supsl∈(0,1)
(sl
1− (1− 3ωJac2 sl)ν
+1− sl
1− (1− 3ωJac2 (1− sl))ν
)=
23νωJac
+1
1− (1− 3ωJac2 )ν
,
with ωJac ∈ ( 0 , 4/3 ). More precisely, we prove that
f(sl) =sl
1− (1− 3ωJac2 sl)ν
is a convex function for ωJac ∈ ( 0 , 4/3 ), and hence so is f(sl) + f(1 − sl), the prove
being finished by the fact that any convex function takes it supremum at the boundary.
Now, note that
f(c) =3ωJac
2f(c (2/3)ω−1
Jac) =(1 + c+ ...+ cν−1
)−1 = g(c)−1
is convex for c ∈ (−1, 1) if and only if f(sl) is convex. However, f(c) is convex if d2fdc2
> 0
for c ∈ (−1, 1), that is, if d2gdc2· g < 2 ·
(dgdc
)2. On the other hand, one can check that
d2g
dc2· g − 2
(dg
dc
)2
= −ν/2−1∑i=0
c2i−2(c2 + i(ν − 2i)(c+ 1)2) ,
this last term being negative for c ∈ (−1, 1).
Chapter 3When does two-grid optimality carry over to the
V-cycle?
Summary
We investigate additional condition(s) that confirm that a V-cycle multigrid method is
satisfactory (say, optimal) when it is based on a two-grid cycle with satisfactory (say,
level-independent) convergence properties. The main tool is McCormick’s bound on
the convergence factor [SIAM J. Numer.Anal., 22(1985), pp.634-643], which we showed
in previous work to be the best bound for V-cycle multigrid among those that are
characterized by a constant that is the maximum (or minimum) over all levels of an
expression involving only two consecutive levels; that is, that can be assessed considering
only two levels at a time. We show that, given a satisfactorily converging two-grid
method, McCormick’s bound allows us to prove satisfactory convergence for the V-
cycle if and only if the norm of a given projector is bounded at each level. Moreover,
this projector norm is simple to estimate within the framework of Fourier analysis,
making it easy to supplement a standard two-grid analysis with an assessment of the
V-cycle potentialities. The theory is illustrated with a few examples that also show that
the provided bounds may give a satisfactory sharp prediction of the actual multigrid
convergence.
3.1 Introduction
We consider multigrid methods for the solution of symmetric positive definite (SPD)
n× n linear systems:
Ax = b. (3.1)
Multigrid methods are based on the recursive use of a two–grid scheme. A basic two–
grid method combines the action of a smoother, often a simple iterative method such
as Gauss-Seidel, and a coarse-grid correction, which corresponds to the solution of the
37
38 When does two-grid optimality carry over to the V-cycle?
residual equation on a coarser grid. A V–cycle multigrid method is obtained when the
residual equation is solved approximately with one application of the two–grid scheme
on that level, and so on, until the coarsest level, where an exact solve is performed.
Other cycles may be defined, including the W–cycle based on two recursive applications
of the two-grid scheme on each level, see, e.g., [61].
If there are only two levels, accurate bounds may be obtained either by means of
Fourier analysis [60,61,68], or by using some appropriate algebraic tools [16,22,23,46,59].
This focus on two-grid schemes is motivated by the fact that, “if the two-grid method
converges sufficiently well, then the multigrid method with W–cycle will have similar
convergence properties” [61, p. 77] (see also [12, pp. 226–228] and [47]). This is not
the case for the V–cycle since there are known examples where the two-grid method
converges relatively well, whereas the multigrid method with V–cycle scales poorly with
the number of levels [41]. Hence, V–cycle analysis has to be, at some point, essentially
different from two-grid analysis.
In this chapter, we investigate additional condition(s) for obtaining an optimal V-
cycle method from an optimal1 two-grid method. Note that we do not base our work
on a new analysis of the V-cycle. Several analyses are indeed available, which, however,
have a common gap: the conditions for proving that the V-cycle converges nicely have
not been compared with the two-grid convergence factor, and it is so far unclear how
they are related. In fact, a number of results relate the V-cycle convergence to sufficient
conditions for two-grid convergence; see, e.g., the two conditions (3.3) in [14], the first of
which is sufficient for two-grid. Or, simply, consider V-cycle analysis particularized to
the two-level case. Such sufficient conditions are, however, often stronger than needed
for just two-level convergence, and, as far as we know, no comparison has been made
with necessary and sufficient conditions or with two-grid convergence factor.
To analyze the V-cycle, one possibility consists of defining an appropriate sub-
space decomposition and then applying successive subspace correction (SSC) theory
[50, 51, 25, 73, 75, 74]. Another possibility consists in checking so-called smoothing and
approximation properties [10, 13, 26, 27, 37, 38, 53]. Regarding the latter approach, the
best result for SPD matrices have been obtained by Hackbusch [27, Theorem 7.2.2] and
McCormick [38]. In a Chapter 2, we show that these results are qualitatively equiv-
alent, with McCormick’s bound being always the sharpest. Note that, in both cases,
the bound is characterized by a constant that is the minimum/maximum over all levels
of an expression involving only two consecutive levels. This last property is important
in the context of this study, since it seems at first sight not possible to compare with1By “optimal”, for a two-grid method, we mean “having level-independent convergence properties”;
that is, referring to a situation where the two-grid method is defined at different levels of a multigridhierarchy, it is considered optimal if there is a level-independent bound on the convergence factor thatis uniform with respect to problem size.
When does two-grid optimality carry over to the V-cycle? 39
the two-grid convergence rate a global expression that would involve simultaneously all
levels.
On the other hand, we also consider in Chapter 2 the classical formulation of the
SSC theory (as stated in [73] or [75]), and discuss how to obtain a bound that could
also be assessed considering only two levels at a time. It turns out that this requires the
use of the so-called a-orthogonal decomposition, which corresponds to the choice most
frequently made when applying the SSC theory to multigrid methods for H2-regular
problems. Then, the analysis in Chapter 2 shows that this approach is also qualitatively
equivalent to the Hackbusch and McCormick ones, the latter remaining the sharpest.
Hence, regarding the goal pursued in this work, all exploitable results are superseded
by (but qualitatively equivalent to) McCormick’s bound, which is characterized by the
constant δ; in this work, we relate this constant to the two-grid convergence factor. This
reveals that a satisfactory (optimal) two-grid cycle on each level leads to a satisfactory
estimate of δ if and only if a given norm of an exact coarse-grid correction (projection)
operator remains bounded at each level. Moreover, it turns out that this norm is easy
to assess within the framework of a Fourier analysis.
Eventually, we consider several examples, illustrating the sharpness of the bound
based on two-grid convergence rates and the projector norm. It further turns out that
both of these ingredients are independent and play an important role in the V-cycle
convergence behavior.
The reminder of this chapter is organized as follows. In Section 3.2 we state the gen-
eral setting of this study and gather the needed assumptions. The relation between the
McCormick constant δ and the two-grid convergence factor is established in Section 3.3.
Illustrative examples are discussed in Section 3.4.
3.2 General setting
We consider a multigrid method with J + 1 levels (J ≥ 1); index J refers to the finest
level (on which the system (3.1) is to be solved), and index 0 to the coarsest level. The
number of unknowns at level k , 0 ≤ k ≤ J , is noted nk (with thus nJ = n).
Our analysis applies to symmetric multigrid schemes based on the Galerkin principle
for the SPD system (3.1); that is, restriction is the transpose of prolongation and the
matrix Ak at level k , k = J − 1, . . . , 0 , is given by Ak = P Tk Ak+1Pk , where Pk is the
prolongation operator from level k to level k+ 1 ; we also assume that the smoother Rkis SPD and that the number of pre–smoothing steps ν (ν > 0) is equal to the number
of post–smoothing steps. The algorithm for V–cycle multigrid is then as follows.
40 When does two-grid optimality carry over to the V-cycle?
Multigrid with V–cycle at level k: xn+1 = MG(b, Ak,xn, k)
(1) Relax ν times with smoother Rk : xn ← Smooth(xn, Ak, Rk, ν,b)
(2) Compute residual: rk = b−Akxn(3) Restrict residual: rk−1 = P Tk−1rk(4) Coarse grid correction: if k = 1 , e0 = A−1
0 r0
else ek−1 = MG(rk−1, Ak−1, 0, k − 1)(5) Prolongate coarse-grid correction: xn ← xn + Pk−1ek−1
(6) Relax ν times with smoother Rk : xn+1 ← Smooth(xn, Ak, Rk, ν,b)
When applying this algorithm, the error satisfies
A−1k b− xn+1 = E
(k)MG
(A−1k b− xn
),
where the iteration matrix E(k)MG is recursively defined from
E(0)MG = 0 and, for k = 1, 2, . . . , J :
E(k)MG = (I −R−1
k Ak)ν(I − Pk−1(I − E(k−1)
MG )A−1k−1P
Tk−1Ak
)(I −R−1
k Ak)ν(3.2)
(see, e.g., [61, p. 48]). Our main objective is the analysis of the spectral radius of E(J)MG ,
which governs convergence on the finest level. Our analysis makes use of the following
general assumptions.
General assumptions
• n = nJ > nJ−1 > ... > n0 ;
• Pk is an nk+1 × nk matrix of rank nk, k = J − 1, . . . , 0 ;
• AJ = A and Ak = P Tk Ak+1Pk , k = J − 1, . . . , 0 ;
• Rk is SPD and such that ρ(I −R−1k Ak) < 1 , k = J, . . . , 1 .
In what follows, we make use of the two-grid cycle involving two consecutive levels k
and k − 1, which corresponds to the following iteration matrix:
E(k)TG = (I −R−1
k Ak)ν(I − Pk−1A
−1k−1P
Tk−1Ak
)(I −R−1
k Ak)ν , k = 1, . . . , J . (3.3)
Most of our results do not refer explicitly to the smoother Rk , but are stated with
respect to the matrices M (ν)k defined from
I − M(ν)k
−1Ak = (I −R−1
k Ak)ν . (3.4)
That is, M (ν)k is the smoother that provides in one step the same effect as ν steps with
Rk . The results stated with respect to M(ν)k may then be seen as results stated for the
When does two-grid optimality carry over to the V-cycle? 41
case of one pre– and one post–smoothing step, which can be extended to the general
case via the relations (3.4).
We close this subsection by introducing the projector πAk , which plays an important
role throughout this chapter:
πAk = Pk−1A−1k−1P
Tk−1Ak . (3.5)
Note that I − πAk is the (exact) coarse-grid correction matrix at level k.
3.3 Theoretical Analysis
3.3.1 McCormick’s bound
We recall in the following theorem the bound obtained in [38, Lemma 2.3, Theorem 3.4
and Section 5] (see also [37], or [53] for an alternative proof). The equivalence of (3.8)
with the definition (3.7) is proved in in Theorem 2.6.
Note that convergence estimates based on regularity assumptions are also considered
in [37]. These estimates are obtained when Theorem 3.1 below is applied to discretized
PDEs. However, Theorem 3.1 on its own is a purely algebraic result that may by applied
to any multigrid method satisfying the general assumptions in Section 3.2, without
reference to a PDE context. Hence, there is no need for regularity assumptions to apply
here, as may be further confirmed by the purely algebraic proof in [53].
Theorem 3.1. Let E(J)MG , M (ν)
k , and πAk , k = 1, . . . , J , be defined, respectively, by
(3.2), (3.4), and (3.5), with Pk , k = 0, . . . , J − 1 , Ak , k = 0, . . . , J , and Rk , k =
1, . . . , J , satisfying the general assumptions stated in Section 3.2.
Then
ρ(E(J)MG) ≤ 1− δ(ν) , (3.6)
where
δ(ν) = min1≤k≤J
minvk∈Rnk
‖vk‖2Ak − ‖(I − M(ν)k
−1Ak)vk‖2Ak
‖(I − πAk)vk‖2Ak(3.7)
= min1≤k≤J
minvk∈Rnk
vTkAkvk
vTk (I − πAk)TM (2ν)k (I − πAk)vk
(3.8)
3.3.2 Relationship to the two-grid convergence rate
We first recall, in the following lemma, a useful characterization of the two-grid rate
obtained in [23, p. 480].
42 When does two-grid optimality carry over to the V-cycle?
Lemma 3.1. Let E(k)TG , M (ν)
k , and πAk , k = 1, . . . , J , be defined, respectively, by (3.3),
(3.4), and (3.5), with Pk , k = 0, . . . , J − 1 , Ak , k = 0, . . . , J , and Rk , k = 1, . . . , J ,
satisfying the general assumptions stated in Section 3.2.
Then
1− ρ(E(k)TG) = min
vk∈Rnk
vTk (I − πAk)A1/2k M
(2ν)k
−1A
1/2k (I − πAk)vk
vTk (I − πAk)vk, (3.9)
with πAk = A1/2k πAkA
−1/2k .
The next theorem contains our main result.
Theorem 3.2. Let E(k)TG , M (ν)
k , and πAk , k = 1, . . . , J , be defined, respectively, by
(3.3), (3.4), and (3.5), with Pk , k = 0, . . . , J − 1 , Ak , k = 0, . . . , J , and Rk , k =
1, . . . , J , satisfying the general assumptions stated in Section 3.2. Let δ(ν) be defined by
(3.7).
Then
δ(ν) ≥ min1≤k≤J
1− ρ(E(k)TG)
‖I − πAk‖2M(2ν)k
= min1≤k≤J
1− ρ(E(k)TG)
‖πAk‖2M(2ν)k
. (3.10)
Moreover,
δ(ν) ≤ min1≤k≤J
min
1− ρ(E(k)TG) ,
1‖πAk‖2M(2ν)
k
. (3.11)
Proof.
Let ξk be defined by
ξk = minv∈Rnk
vTAkv
vT (I − πAk)TM (2ν)k (I − πAk)v
.
From (3.8), there holds
δ(ν) = min1≤k≤J
ξk . (3.12)
On the other hand, Lemma 3.1 implies (since Ak(I−πAk) = (I−πAk)TAk and (I−πAk) =
(I − πAk)2)
1− ρ(E(k)TG) = min
vk∈Rnk
vTkA1/2k (I − πAk) M (2ν)
k
−1Ak(I − πAk)A−1/2
k vk
vTkA1/2k (I − πAk)A−1/2
k vk
= minvk∈Rnk
vTk (I − πAk) M (2ν)k
−1Ak(I − πAk)A−1
k vkvTk (I − πAk)(I − πAk)A−1
k vk
= minvk∈Rnk
vTk (I − πAk) M (2ν)k
−1(I − πAk)Tvk
vTk (I − πAk)A−1k (I − πAk)Tvk
. (3.13)
When does two-grid optimality carry over to the V-cycle? 43
In what follows, we omit the subscripts k, as well as the superscript (k) and (2ν)in ETG and M , respectively, when they are obvious from context. Using (3.13), oneobtains
ξ−1 = maxv∈Rn
vT (I − πA)TM(I − πA)vvTAv
= maxv∈Rn
vTA−1/2(I − πA)TM1/2M1/2(I − πA)A−1/2vvTv
= maxv∈Rn
vTM1/2(I − πA)A−1/2A−1/2(I − πA)TM1/2vvTv
= maxv∈Rn
vT (I − πA)A−1(I − πA)TvvTM−1v
(3.14)
≤ maxv∈Rn
vT (I − πA)A−1(I − πA)TvvT (I − πA)M−1(I − πA)Tv
maxv∈Rn
vT (I − πA)M−1(I − πA)TvvTM−1v
=1
1− ρ(ETG)maxv∈Rn
vTM1/2(I − πA)M−1/2M−1/2(I − πA)TM1/2vvTv
=1
1− ρ(ETG)maxv∈Rn
vTM−1/2(I − πA)TM1/2M1/2(I − πA)M−1/2vvTv
=1
1− ρ(ETG)maxv∈Rn
vT (I − πA)TM(I − πA)vvTMv
=1
1− ρ(ETG)‖I − πA‖2M .
The result (3.10) follows directly, using Kato’s lemma (e.g., [65, Lemma 3.6]) which
implies ‖I − πA‖M = ‖πA‖M , since πA 6= O, I by virtue of our general assumptions.
In addition, using (3.14) together with Lemma 3.1, one also has
ξ = minv∈Rn
vTM−1vvT (I − πA)A−1(I − πA)Tv
≤ minv=(I−πA)Tw, w∈Rn
vTM−1vvT (I − πA)A−1(I − πA)Tv
= 1− ρ(ETG),
which gives the first term in the right-hand side of (3.11).
On the other hand, since
vTA1/2 M (2ν)−1A1/2v = vTv − vT (I −A1/2 M (ν)−1
A1/2)2vT ≤ vTv , ∀v ∈ Rn ,
there holds
vTAv ≤ vTMv , ∀v ∈ Rn .
44 When does two-grid optimality carry over to the V-cycle?
Hence,
ξ = minv∈Rn
vTAvvT (I − πA)TM(I − πA)v
≤ minv∈Rn
vTMvvT (I − πA)TM(I − πA)v
=1
‖I − πA‖2M,
(3.15)
which, combined with Kato’s lemma ‖I − πA‖M = ‖πA‖M , gives the second term in the
right-hand side of (3.11).
Theorem 3.2 shows that McCormick’s bound proves a satisfactory convergence rate
for the V–cycle if and only if, at each level, the two-grid method converges fast enough
and ‖πAk‖M(2ν)k
= ‖M (2ν)k
1/2πAk M
(2ν)k
−1/2‖ is nicely bounded. We can further show
the following corollary.
Corollary 3.1. Let the assumptions of Theorem 3.2 hold and let E(J)MG be defined by
(3.2).
Then
ρ(E(J)TG) ≤ ρ(E(J)
MG) ≤ 1− δ(ν) ≤ 1− min1≤k≤J
1− ρ(E(k)TG)
‖πAk‖2M(2ν)k
. (3.16)
Proof.
The proof of ρ(E(k)TG) ≤ ρ(E(k)
MG) can be deduced from the relation (7.2.2a) in [27] com-
bined with (7.2.4a) from the same reference, which proves that
A1/2E(k)MGA
−1/2 ≤ A1/2E(k)TGA
−1/2 .
The other results follow from Theorems 3.1 and 3.2.
Note that the V-cycle convergence factor is bounded below by the two-grid conver-
gence factor on the finest grid only. Indeed, max1≤k≤J ρ(E(k)TG) can be close to 1 even
when ρ(E(J)MG) is not, for instance when the smoother alone is efficient enough on the
finest level, so that poor two-grid ingredients on coarser levels will not significantly affect
the convergence. In practice, however, one has often max1≤k≤J ρ(E(k)TG) ≈ ρ(E(J)
TG) (e.g.,
consider the discrete Poisson equation on many simple geometries with uniform meshes).
Then (3.16) defines an interval, containing both 1− δ(ν) and ρ(E(J)MG), that is narrow if
and only if max1≤k≤J ‖πAk‖M(2ν)k
is not much larger than 1.
3.3.3 Fourier analysis
Often, a multigrid method is assessed by estimating the two-grid convergence rate with
Fourier analysis [60,61,68]. This means that one considers a model constant-coefficient
When does two-grid optimality carry over to the V-cycle? 45
PDE for which the eigenvectors of the discrete matrix are explicitly known at all levels.
Simple smoothers have the same set of eigenvectors and, hence, the matrices Ak and
Rk are both diagonal whenever expressed in the corresponding basis (the Fourier basis).
In more complicated situations, Rk may be only block-diagonal with small diagonal
blocks; Ak may also have a block diagonal structure in case of coupled systems of PDEs.
Note that M (2ν)k , expressed in the Fourier basis, will then have the same block diagonal
structure as Ak and Rk, and will be pointwise diagonal if Ak and Rk are pointwise
diagonal.
Let
Ak =
Λ(k)
1
Λ(k)2
. . .
Λ(k)lk
, M(2ν)k =
Σ(k)
1
Σ(k)2
. . .
Σ(k)lk
be this (block) diagonal representation of Ak and M
(2ν)k , where the ith block has size
m(k)i ×m
(k)i , i = 1, ..., lk. Technically, Fourier analysis of a two-grid method at level k
characterized by a given prolongation Pk−1 is possible if there exists a basis of the coarse
space (the coarse Fourier basis) such that the expression of Pk−1 in both this basis and
the (fine grid) Fourier basis has the structure
Pk−1 =
p
(k−1)1
p(k−1)2
. . .
p(k−1)lk
,
where p(k−1)i are (possibly complex) rectangular matrices of size m(k)
i ×m(k−1)i .
Here, we observe that, in this context, M (2ν)k
1/2πAk M
(2ν)k
−1/2is also block diagonal
with diagonal blocks of the form
Σ(k)i
1/2p
(k−1)i
(p
(k−1)i
HΛ(k)i p
(k−1)i
)−1
p(k−1)i
HΛ(k)i Σ(k)
i
−1/2. (3.17)
Hence, ‖πAk‖2M(2ν)k
is the maximal norm of all these m(k)i × m
(k)i blocks. Further, the
matrices (3.17) are the product of rectangular matrices; taking the product of theirnorms gives an easy-to-assess upper bound:
‖πAk‖M(2ν)k
≤ maxi
∥∥ Σ(k)i
1/2p(k−1)i
∥∥ ∥∥(p(k−1)i
HΛ(k)i p
(k−1)i
)−1
p(k−1)i
HΛ(k)i Σ(k)
i
−1/2 ∥∥. (3.18)
It is worth noting that the latter inequality becomes an equality when m(k−1)i = 1 for
46 When does two-grid optimality carry over to the V-cycle?
all i; that is, when the rectangular blocks p(k−1)i are all simple vectors, as most often
arises when analyzing scalar PDEs.
3.3.4 Finite element setting
Consider a finite element discretization of Poisson boundary value problem on a bounded
domain. Such a domain is first approximated by an appropriate polygonal or polyhedral
mesh, which is then refined several times. These refinements naturally induce a multigrid
hierarchy (including inter-grid transfer operators Pk). It then can be shown (see [72,
Theorem 4.2]) that ‖πAk‖ are bounded on all levels if and only if the underlying problem
possesses (full) elliptic regularity. Since ‖ · ‖ behaves similarly to ‖ · ‖M
(2ν)k
for a number
of smoothers, essentially the same conclusions hold with respect to ‖πAk‖M(2ν)k
.
With regards to the Theorem 3.2, these observations show that level independent
two-grid convergence implies, in this context, a level-independent bound for V-cycle
multigrid if and only if the problem has full elliptic regularity. Hence, it follows that
McCormick’s analysis cannot prove optimal bounds for the V-cycle if the problem does
not possess full regularity. Considering the results in Chapter 2, the same conclusions
hold for Hackbusch’s analysis [27, Section 7.2], and the successive subspace correction
theory with a-orthogonal decomposition [73, 75]. Thus, for the case when ‖πAk‖ and
‖πAk‖M(2ν)k
behave similarly with respect to the problem size, we show here that another
type of analysis, as developed in, e.g., [50,51,25,73,75,74], is really needed to get uniform
results for the V-cycle for problems with less than full regularity.
3.4 Examples
We consider three examples that represent three possible different practical situations.
In the first, both ρ(E(k)TG) and ‖πA‖2M(2) are nicely bounded above. In the second exam-
ple, ρ(E(k)TG) remains bounded away from one while ‖πA‖2M(2) increases rapidly with the
problem size. The third example is the other way around: ‖πA‖2M(2) is nicely bounded
while ρ(E(k)TG) is far from being optimal.
3.4.1 Standard multigrid with 2D Poisson
We consider the linear system resulting from the bilinear finite element discretization of
the two-dimensional Poisson problem
−∆u = f in Ω = (0, 1)× (0, 1)
u = 0 in ∂Ω
When does two-grid optimality carry over to the V-cycle? 47
on a uniform grid of mesh size h = 1/NJ in both directions. The matrix corresponds
then to the following nine point stencil:−1 −1 −1
−1 8 −1
−1 −1 −1
. (3.19)
Up to some scaling factor, this is also the stencil obtained with 9-point finite difference
discretization. We assume NJ = 2JN0 for some integer N0 , allowing J steps of regular
geometric coarsening. We consider the standard prolongation operator
Pk =
(Jk
Ink
),
where Jk corresponds to the natural interpolation associated with bilinear finite element
basis functions. The restriction P Tk corresponds then to “full weighting”, as defined
in, e.g., [61] 2. We consider damped Jacobi smoothing: Rk = ω−1Jacdiag(Ak) . Since the
stencil is preserved on all levels, it is sufficient to consider only two successive grids; to
alleviate notation, we therefore let N = Nk , A = Ak ,R = Rk , M = M(ν)k , P = Pk−1 ,
Ac = Ak−1 = P TAP , and πA = πAk = PA−1c P TA .
We now use Fourier analysis to asses ‖πA‖M(2ν) via (3.18). The eigenvectors of A
are, for i, j = 1, . . . , N − 1 , the functions
u(N)i,j = sin(iπx) sin(jπy)
evaluated at the grid points. The eigenvalue corresponding to u(N)i,j is
λ(N)i,j = 4(3si + 3sj − 4sisj) , (3.20)
where
si = sin2(iπ
2N) , sj = sin2(
jπ
2N) . (3.21)
Hence, the eigenvalues of I −R−1A are in the interval [1− ωJac 32 , 1). One has therefore
ρ(I − R−1A) ≤ 1, as required by our general assumptions if ωJac ∈ (0, 4/3). The
prolongation P satisfies (see, e.g., [61, p. 87])
P T
u
(N)i,j
u(N)N−i,N−j
−u(N)N−i,j
−u(N)i,N−j
= 4
(1− si)(1− sj)
sisj
si(1− sj)(1− si)sj
u
(N/2)i,j
2up to some scaling factor; the scalings of the prolongation and restriction are unimportant whenusing coarse-grid matrices of the Galerkin type.
48 When does two-grid optimality carry over to the V-cycle?
for 1 ≤ i, j ≤ N/2− 1 , with P Tu(N)i,j = 0 for i = N/2 or j = N/2 . Using
pi,j = 4(
(1− si)(1− sj) sisj si(1− sj) (1− si)sj)T
,
Λi,j = diag(λ
(N)i,j , λ
(N)N−i,N−j , λ
(N)N−i,j , λ
(N)i,N−j
),
Σ(ν)i,j = diag
σ(ν)(λ(N)c,s )
∣∣∣∣∣ σ(ν)(λ) =λ
1− (1− ωJacλ8 )ν
(c,s)=(i,j),(N−i,N−j),(N−i,j),(i,N−j)
,
we can rewrite (3.18):
‖πA‖2M(2ν) = maxi,j=1,...,N−1
g(ν)(si, sj) ,
where
g(ν)(si, sj) =
∥∥ Σ(2ν)i,j
1/2pi,j∥∥2 ∥∥ pi,j T Λi,j Σ(2ν)
i,j
−1/2 ∥∥2
(pi,j T Λi,j pi,j)2 . (3.22)
One also has
maxi,j=1,...,N−1
g(ν)(si, sj) ≤ sup(si,sj)∈(0,1)×(0,1)
g(ν)(si, sj) .
For all ν, g(ν)(si, sm) exhibits the following symmetries: g(ν)(si, sj) = g(ν)(1−si, sj) =
g(ν)(si, 1−sj) = g(ν)(1−si, 1−sj) . Further, numerical investigations reveal that the
maximum on the considered domain is located at the boundary, i.e., corresponds to,
e.g., sj = 0 or, equivalently, j = 0 (such index values represent asymptotic behavior and
do not correspond to any Fourier block) . Because of the symmetries, it is sufficient to
analyze this latter case. Next, since
g(ν)(si, 0)
=
((pi,0)2
1σ(2ν)(λ(N)
i,0 ) + (pi,0)23σ
(2ν)(λ(N)N−i,0)
)( (pi,0)21
(λ(N)i,0
)2
σ(2ν)(λ(N)i,0 )
+(pi,0)23
(λ(N)N−i,0
)2
σ(2ν)(λ(N)N−i,0)
)(
(pi,0)21λ
(N)i,0 + (pi,0)2
3λ(N)N−i,0
)2
= 1 +(pi,0)2
1(pi,0)23
(σ(2ν)(λ
(N)i,0 )
σ(2ν)(λ(N)N−i,0)
(λ
(N)N−i,0
)2+
σ(2ν)(λ(N)N−i,0)
σ(2ν)(λ(N)i,0 )
(λ
(N)i,0
)2− 2λ(N)
i,0 λ(N)N−i,0
)(
(pi,0)21λi,0 + (pi,0)2
3λN−i,0)2
= 1 + si(1−si)
(1−(1− 3
2ωJacsi)2ν
1−(1− 3
2ωJac(1−si))2ν +
1−(1− 3
2ωJac(1−si))2ν
1−(1− 3
2ωJacsi)2ν −2
), (3.23)
When does two-grid optimality carry over to the V-cycle? 49
ωJac 1− 1‖πA‖2
M(2)
ρ(E(J)TG) ρ(E(J)
MG) 1− δ(1) 1− (1−ρ(E(J)TG))
‖πA‖2M(2)
1/2 0.385 0.391 0.398 0.423 0.6252/3 0.333 0.25 0.271 0.333 0.51 0.2 0.25 0.251 0.4 0.4
Table 3.1: The estimates of main convergence parameters for ν = 1 and for differentdamping factors ωJac.
ωJac 1− 1‖πA‖2
M(4)
ρ(E(J)TG) ρ(E(J)
MG) 1− δ(2) 1− (1−ρ(E(J)TG))
‖πA‖2M(4)
1/2 0.25 0.153 0.187 0.252 0.3652/3 0.2 0.083 0.121 0.2 0.2661 0.143 0.068 0.091 0.189 0.2
Table 3.2: The estimates of main convergence parameters for ν = 2 and for differentdamping factors ωJac.
we obtain (see Appendix A for details)
‖πA‖2M(2ν) ≤ sup(si,sj)∈(0,1)×(0,1)
g(ν)(si, sj) = supsi∈(0,1)
g(ν)(si, 0) ≤
2− 3ωJac4 if ν = 1
1 + 13νωJac
if ν > 1 .
Note that this bound is asymptotically sharp for N → ∞ when ν = 1, since
lims→0 g(1)(s, 0) = 2 − 3ωJac/4. In Tables 3.1 and 3.2, we use this bound and the
asymptotically sharp estimate
δ(ν)−1 ≤ 13νωJac
+1
1− (1− 3ωJac2 )2ν
, ∀ ν = 1, 2,
obtained in Chapter 2 to illustrate inequalities (3.16), with two-grid and V-cycle multi-
grid convergence factors numerically assessed for N0 = 2 and J = 7 (hence N = 256).
Note that ρ(E(k)TG) increases with the mesh size, so that max1≤k≤J ρ(E(k)
TG) corresponds
to the value on the finest grid, which is close to the asymptotic one. Observe that the
interval containing both ρ(E(J)MG) and 1 − δ(1) is sharp enough. On the other hand,
1− 1‖πA‖2
M(2)
is also a lower bound on 1− δ(1) by (3.11), but in general not a lower bound
on the effective convergence factor.
50 When does two-grid optimality carry over to the V-cycle?
3.4.2 Aggregation-based multigrid for 1D Poisson
We consider N ×N linear system associated to A = A(ε), where
A(ε) =
2 −1 · · · −1
−1 2 −1
−1 2. . .
......
. . . . . . −1
−1 · · · −1 2
+ ε N−1 IN , (3.24)
with N = 2JN0 and ε > 0. We also assume piecewise constant prolongation of the form
P =
1 1
1 1. . .
1 1
T
.
Note that, with this prolongation, the successive coarse-grid matrices Ak = Ak(ε) are also
given by (3.24) withN replaced byNk = 2kN0, where we considerN0 ≥ 2. Hence, we can
omit the subscript k (or k−1), let Ac = Ak−1 = P TAP , and set πA = πAk = PA−1c P TA.
Note that this is a 1D like problem which could be solved more efficiently using a tri-
diagonal solver. The analysis below can however be easily repeated in more dimensions,
leading essentially to the same conclusions. We therefore continue with the 1D variant
for the sake of simplicity.
The eigenvectors of A(ε) are, for j = 0, . . . , N − 1 , the functions
u(N)j =
1√N
exp(i jπx)
evaluated at the grid points, with i =√−1. The eigenvalue corresponding to u(N)
j is
λ(N)j (ε) = 4 sin2(jπN−1) + ε N−1 .
The prolongation P satisfies (see [41, p. 1087])
P T
u
(N)j
u(N)j+N/2
=√
2 ei jπN−1
cos(jπN−1)
i sin(jπN−1)
u
(N/2)j .
We consider damped Jacobi smoother R = 2 diag(A). Hence, the eigenvalues of
I − R−1A are in the interval [1 − ε N−1
4+2ε N−1 , 1 − 4+ε N−1
4+2ε N−1 ) = [ω, 1 − ω) with ω =4+ε N−1
4+2ε N−1 ∈ (0, 1). One therefore has ρ(I − R−1A) ≤ 1, as required by our general
assumptions.
When does two-grid optimality carry over to the V-cycle? 51
Letting
pj =√
2 ei jπN−1(
cos(jπN−1) i sin(jπN−1))H
,
Λj(ε) = diag(λ
(N)j (ε) , λ(N)
j+N/2(ε)),
Σ(ν)j (ε) = diag
σ(ν)(λ(N)c (ε))
∣∣∣∣∣ σ(ν)(λ) =λ
1− (1− ωλ4+ε N−1 )ν
(c)=(j),(j+N/2)
,
we can rewrite (3.18):
‖πA‖M(2ν) = maxj=0,...,N/2−1
∥∥ Σ(2ν)j (ε)
1/2pj∥∥ ∥∥ pj H Λj(ε) Σ(2ν)
j (ε)−1/2 ∥∥
pj H Λj(ε) pj. (3.25)
First observe that σ(2ν)(λ) is an increasing function of λ since t(1 − (1 − t)2ν)−1 isan increasing function of t on the interval (0, 1). Hence, since λ(N)
1 (ε) ≤ λ(N)1+N/2(ε) for
N ≥ 2N0 ≥ 4, we have
‖πA‖M(2ν) ≥∥∥ Σ(2ν)
1 (ε)1/2
p1
∥∥ ∥∥ p1H Λ1(ε) Σ(2ν)
1 (ε)−1/2 ∥∥
p1H Λ1(ε) p1
=
√|(p1)1|2 σ(2ν)(λ
(N)1 (ε))
σ(2ν)(λ(N)1+N/2(ε))
+ |(p1)2|2√|(p1)1|2 λ(N)
1 (ε)2 σ(2ν)(λ
(N)1+N/2(ε))
σ(2ν)(λ(N)1 (ε))
+ |(p1)2|2 λ(N)1+N/2(ε)
2
|(p1)1|2 λ(N)1 (ε) + |(p1)2|2 λ(N)
1+N/2(ε)
≥
√√√√ σ(2ν)(λ(N)1 (ε))
σ(2ν)(λ(N)1+N/2(ε))
√|(p1)1|2 + |(p1)2|2
√|(p1)1|2 λ(N)
1 (ε)2
+ |(p1)2|2 λ(N)1+N/2(ε)
2
|(p1)1|2 λ(N)1 (ε) + |(p1)2|2 λ(N)
1+N/2(ε)
=
√√√√ σ(2ν)(λ(N)1 (ε))
σ(2ν)(λ(N)1+N/2(ε))
√cos4(πN−1) sin2(πN−1) + cos2(πN−1) sin4(πN−1) +O(ε)
2 cos2(πN−1) sin2(πN−1) +O(ε).
Further, using again the monotonicity of σ(2ν), there holds
σ(2ν)(λ(N)1 (ε))
σ(2ν)(λ(N)1+N/2(ε))
≥ limλ→0 σ(2ν)(λ)
σ(2ν)(4 + εN−1)=
(4 + εN−1)νω
1− (1− ω)2ν
(4 + εN−1)=
1− (1− ω)2ν
νω
with ω ∈ (0, 1). Hence, for ε→ 0, we have
‖πA‖2M(2ν) ≥1− (1− ω)2ν
νω
14 cos2(πN−1) sin2(πN−1)
= O(N2) .
Thus, ‖πA‖2M
(2)k
increases with the problem size when ε is small enough, whereas, as
shown in [41], the two-grid convergence factor remains bounded. Hence, we have an
example of optimal two-grid method for which the V-cycle convergence estimate is poor.
As seen in Table 3.3, it turns out that the actual convergence factor also deteriorates with
the number of levels, showing that the analysis based on ‖πA‖2M(2) is qualitatively correct.
52 When does two-grid optimality carry over to the V-cycle?
J(N) 1(8) 3(32) 5(128) 7(512) 9(2048)‖πA‖2
M(2)J
1.471 13.58 208.0 3312 52575
ρ(E(J)TG) 0.375 0.490 0.499 0.5 0.5
ρ(E(J)MG) 0.375 0.800 0.947 0.986 0.997
Table 3.3: The values of main parameters for ε = 10−4 and for different problemsizes; the coarsest grid corresponds to N0 = 4.
3.4.3 Positive off-diagonal entries
We consider the (2NJ − 1)× (2NJ − 1) matrix
A =
2 1
1 2 1. . . . . . . . .
1 2 1
1 2
,
with Nk = N0 · 2k, corresponding to the one-dimensional stencil[1 2 1
]. (3.26)
We also consider the (2Nk − 1)× (Nk − 1) prolongation matrix
Pk =1√2
1 0 1
1 0 1. . .
1 0 1
T
(3.27)
and the damped Jacobi smoother Rk = 12 diag(Ak) with one pre- and one post-smoothing
step at each level. Note that the stencil (3.26) is preserved on all levels.
The values of ‖πA‖2M
(2)J
and ρ(E(J)TG) on the finest grid, which are also the maximal
values of these parameters over all grids, are given in the Table 3.4 together with the
V-cycle convergence factor ρ(E(J)MG).
J(N) 1(4) 3(16) 5(64) 7(256) 9(1024)‖πA‖2
M(2)J
1.235 1.479 1.498 1.5 1.5
ρ(E(J)TG) 0.625 0.971 0.998 0.9999 0.99999
ρ(E(J)MG) 0.625 0.971 0.998 0.9999 0.99999
Table 3.4: The values of main parameters for different problem sizes; the coarsestgrid corresponds to N0 = 2.
When does two-grid optimality carry over to the V-cycle? 53
This example illustrates that ‖πAk‖2M(2ν)k
is a parameter essentially independent of
ρ(E(k)TG), since it remains nicely bounded while both the two-grid and the V-cycle con-
vergence factor deteriorate rapidly with the problem size.
3.5 Conclusion
We have presented a two-sided inequality (3.16) on the McCormick’s estimate of V-cycle
convergence factor (which is the best bound from the previous chapter). The inequality
proves that the bound predicts an optimal V-cycle convergence if and only if the related
two-grid scheme has level-independent convergence properties and the Mk-norm of a
given projector πAk is bounded on on all levels. As a straightforward consequence, if
the latter norm condition is checked, level-independent two-grid convergence implies
optimal convergence properties for V-cycle multigrid. We have also shown on examples
that both these conditions (level-independent convergence of the two-grid scheme and on
the boundness of the πAk norm) are independent; that is, each of them can be satisfied
whereas the other is not.
In the finite element context, when multigrid hierarchy is induced by successive mesh
refinements, and considering well conditioned smoothers, we have shown that the bound
of McCormick (as well as the other bounds in Chapter 2) provides an optimal estimate for
V-cycle multigrid if and only if the underlying problem possesses (full) elliptic regularity.
Considering the Fourier analysis, we have observed that the norm of πAk can be
easily assessed, allowing to supplement the two-grid estimate with an indication of V-
cycle potentialities.
Appendix A
In this appendix, we outline the proof of the following inequality:
supsi∈(0,1)
g(ν)(si, 0) ≤
2− 3ωJac4 if ν = 1
1 + 13νωJac
if ν > 1 ,(3.28)
with g(ν) defined by (3.23) and ωJac ∈ [ 0 , 4/3 ).
Note that g(ν)(si, 0) = g(ν)(1 − si, 0) and it is sufficient to seek a supremum for
si ∈ (0, 0.5). Next, exchanging (3/2)ωJac for α (hence, α ∈ [0, 2)), one has
g(ν)(si, 0) =1 + si(1− si)([
1− (1− αsi)2ν
1− (1− α(1− si))2ν− 1]
+[
1− (1− α(1− si))2ν
1− (1− αsi)2ν− 1])
=1 + si(1− si)[(1− αsi)2ν − (1− α(1− si))2ν
]×(
11− (1− αsi)2ν
− 11− (1− α(1− si))2ν
)
54 When does two-grid optimality carry over to the V-cycle?
≤1 + si(1− si)[(1− αsi)2ν − (1− α(1− si))2ν
]( 11− (1− αsi)2ν
)(3.29)
≤1 + si(1− αsi)2ν
(1
1− (1− αsi)2ν
)=1 +
(1− αsi)2ν
α∑2ν−1
k=0 (1− αsi)k
≤1 +1
2να,
the last inequality coming from the fact that αsi ∈ [0, 1). This proves (3.28) for ν > 1.
On the other hand, if ν = 1, (3.29) further gives
g(1)(si, 0) ≤1 + si(1− si)[(1− αsi)2 − (1− α(1− si))2
]( 11− (1− αsi)2
)=1 + si(1− si) [α(2− α)(1− 2si)]
(1
αsi(2− αsi)
)=1 + (2− α)(1− 2si)
(1α− 2− αα(2− αsi)
)(3.30)
≤1 + (2− α)(
1α− 2− α
2α
)(3.31)
=2− α
2,
where the inequality (3.31) comes from the fact that the expression (3.30) is a decreasing
function of si. This concludes the proof.
Chapter 4Smoothing factor and actual multigrid
convergence
Summary
We consider the Fourier analysis of multi-grid methods for symmetric positive definite
and semi-positive definite linear systems arising from the discretizations of scalar PDEs.
In this framework, the smoothing factor is frequently used to estimate the potential of
a multigrid approach. In this chapter, the smoothing factor is related to the actual two-
grid convergence rate and also to the V-cycle convergence estimate based on McCormick
theory in [SIAM J. Numer.Anal., 22(1985), pp.634-643]. A two-sided bound is obtained
that defines an interval containing both the two-grid and V-cycle convergence rate.
This interval is narrow when an additional parameter is small enough, which is a simple
function of quantities available in standard Fourier analysis.
From a qualitative viewpoint, it turns out that, besides the smoothing factor, the
convergence mainly depends on the angle between the eigenvectors of the matrix associ-
ated with small eigenvalues and the range of the prolongation. Nice V-cycle convergence
is guaranteed if the tangent of this angle has an upper bound proportional to the eigen-
value, whereas nice two-grid convergence requires the tangent to be bounded by an
expression proportional only to the square root of the eigenvalue.
The presented results apply to rigorous Fourier analysis for regular discrete PDEs,
and also to local Fourier analysis via the discussion of semi-definite systems as may arise
from the discretization of PDEs with periodic boundary conditions.
4.1 Introduction
We consider Fourier analysis of multigrid methods for symmetric positive definite (SPD)
or, more generally, symmetric semi-positive definite n× n linear systems
Ax = b. (4.1)
55
56 Smoothing factor and actual multigrid convergence
Multigrid methods are based on the recursive use of a two–grid scheme. A basic two–
grid method combines the action of a smoother, often a simple iterative method such
as Gauss-Seidel, and a coarse-grid correction, which corresponds to the solution of the
residual equation on a coarser grid. A multigrid method is obtained when the residual
equation is solved approximately applying few iterations of the two–grid scheme on that
level, and so on, until the coarsest level when an exact solve is performed. If the two–grid
method is used recursively once on each level, the resulting algorithm is called V–cycle
multigrid, whereas more involved cycling strategies (like W– or F–cycle) correspond to
more iterations of two–grid method on given levels (see, e.g., [61, 27,67]).
Fourier analysis [60,61,68] is a widely used tool that helps to design efficient multi-
grid approaches. It exploits the fact that the discretization of a constant-coefficient
(elliptic) boundary value problem on simple domains often leads to a system (4.1) of
which discrete Fourier modes are eigenvectors. If, in addition, other multigrid compo-
nents also have a simple block structure in this Fourier basis, the analysis of a multigrid
approach can be reduced to the analysis of diagonal blocks of small size, which can be
done either analytically or numerically. The multigrid components designed for such
simple cases are then adapted to more complex problems.
Fourier analysis is in practice limited to a few consecutive grids: generally two, rarely
three [69]. Often Fourier analysis is further reduced to the computation of a simpler
(one-grid) smoothing factor. When assessing this latter, the coarse-grid correction is
assumed to annihilate the so-called smooth (or low frequency) error modes, while leaving
rough (or high frequency) modes unchanged. Since this is the limit case of the desired
behavior of a coarse-grid correction, the smoothing factor is often considered as an ideal
two-grid convergence estimate. However, it is so far unclear which condition(s) are to be
satisfied by the coarse-grid correction for having the actual two-grid convergence close
to this ideal. Further, nice two-grid convergence does not necessarily imply optimal
convergence of the multigrid method with V-cycle [41], hence the latter likely requires
additional conditions.
In this chapter we investigate these questions for symmetric multigrid schemes of
Galerkin type. The coarse-grid correction is essentially determined by the prolongation,
and we establish a simple connection between the smoothing factor and the actual two-
grid convergence via an additional parameter α that mainly depends on the coefficients
of the prolongation in the Fourier basis. Regarding the V-cycle convergence rate, we
use as main tool McCormick’s bound [38] (see also [37,53]) which is shown in Chapter 2
to be the best convergence estimate among those that can be assessed considering only
two consecutive levels at a time. In a previous chapter, we show that optimal two-grid
convergence implies optimal V-cycle convergence if the norm of the (two-grid) coarse-
grid correction operator is bounded at each level. However, although it is sketched how
to compute this norm within the framework of Fourier analysis, no simple criterion is
Smoothing factor and actual multigrid convergence 57
given nor a connection is made with the smoothing factor. Here we prove a simple
relation between McCormick’s bound and the smoothing factor, using the same easy-to-
compute parameter α that relates the smoothing factor with the two-grid convergence
rate.
When the constant α and the smoothing factor are nicely bounded at each level, our
analysis essentially proves that the two-grid and the V-cycle convergence factors are both
in a narrow interval, which further goes towards zero as the number of smoothing steps
is increased. On the other hand, from a more qualitative viewpoint, we deduce easy-
to-check conditions to be satisfied by the prolongation for optimal two-grid or V-cycle
convergence. Doing so, we give in some sense a more precise meaning to statements like
“Interpolation must be able to approximate an eigenvector with error bound proportional
to the size of the associated eigenvalue” [18, p. 1573], [21, p. 4]. We also highlight that
the conditions for guaranteed optimal V-cycle convergence are in fact stronger that the
conditions for optimal two-grid convergence.
In a number of practical cases, when Fourier analysis cannot be applied directly, it is
still possible to replace boundary conditions, for instance, by the periodic ones, to make
Fourier analysis work. Provided that some negligible extra smoothing is performed
on the boundary, such modification has little influence on the convergence rate [17,
57]. These approaches are closely related to local Fourier analysis, which can often be
viewed [68, Remark 5.3] [61, Section 3.4.4] as a (rigorous) Fourier analysis for problems
with periodic boundary conditions. Since such boundary conditions often lead to semi-
positive definite (singular) systems (4.1), our treatment should be valid for them as
well. This is addressed in this work via the extension of McCormick’s bound to the
semi-positive definite case.
The reminder of this chapter is organized as follows. In Section 4.2 we state the
general setting of this study for SPD systems and gather the needed assumptions. Mc-
Cormick’s bound is introduced in Section 4.3 and Fourier analysis for SPD problems is
discussed in Section 4.4. The approach is extended to symmetric semi-positive definite
systems in Section 4.5. Illustrative examples are discussed in Section 4.6.
4.2 General setting
We consider a multigrid method with J+1 levels; J > 1 corresponds to a truly multigrid
method, whereas J = 1 leads to a mere two-grid scheme. Index J refers to the finest
level (on which the system (4.1) is to be solved), and index 0 to the coarsest level. The
number of unknowns at level k , 0 ≤ k ≤ J , is noted nk (with thus nJ = n).
Our analysis applies to symmetric multigrid schemes based on the Galerkin principle
for the SPD system (4.1); that is, restriction is the transpose of prolongation and the
matrix Ak at level k , k = J − 1, . . . , 0 , is given by Ak = P Tk Ak+1Pk , where Pk is the
58 Smoothing factor and actual multigrid convergence
prolongation operator from level k to level k+ 1 ; we also assume that the smoother Rkis SPD and that the number of pre–smoothing steps ν (ν > 0) is equal to the number
of post–smoothing steps.
The algorithm for V–cycle multigrid is defined as follows.
Multigrid with V–cycle at level k: xn+1 = MG(b, Ak,xn, k)
(1) Relax ν times with smoother Rk :
repeat ν times xn ← xn +R−1k (bk −Akxn)
(2) Compute residual: rk = b−Akxn(3) Restrict residual: rk−1 = P Tk−1rk(4) Coarse grid correction: if k = 1 , e0 = A−1
0 r0
else ek−1 = MG(rk−1, Ak−1,0, k − 1)(5) Prolongate coarse-grid correction: xn ← xn + Pk−1ek−1
(6) Relax ν times with smoother Rk :
repeat ν times xn+1 ← xn +R−1k (bk −Akxn)
Observe that for k = 1 this algorithm corresponds to a standard two-grid method with
exact coarse-grid solve. Our analysis makes use of the following general assumptions.
General assumptions
• n = nJ > nJ−1 > · · · > n0 ;
• Pk is an nk+1 × nk matrix of rank nk, k = J − 1, . . . , 0 ;
• AJ = A and Ak = P Tk Ak+1Pk , k = J − 1, . . . , 0 ;
• Rk is SPD and such that ρ(I −R−1k Ak) < 1 , k = J, . . . , 1 .
Most of our results do not refer explicitly to the smoother Rk , but are stated with
respect to the matrices N (ν)k defined from
N(ν)k =
ν−1∑j=0
(I −R−1k Ak)jR−1
k , (4.2)
which also satisfy
I − N(ν)k Ak = (I −R−1
k Ak)ν . (4.3)
That is, N (ν)k is the relaxation operator that provides in 1 step the same effect as ν steps
with R−1k . The results stated with respect to N (ν)
k may then be seen as results stated for
the case of 1 pre– and 1 post–smoothing step, which can be extended to the general case
via the relation (4.3). If N (ν)k is nonsingular, it plays the same role as M (ν)
k
−1from the
two previous chapters; however, in Section 4.5 potentially singular N (ν)k are considered.
Smoothing factor and actual multigrid convergence 59
When applying the V-cycle algorithm, the error satisfies
A−1k b− xn+1 = E
(k)MG
(A−1k b− xn
)where the iteration matrix E(k)
MG is recursively defined from
E(0)MG = O and, for k = 1, 2, . . . , J :
E(k)MG = (I −R−1
k Ak)ν(I − Pk−1(I − E(k−1)
MG )A−1k−1P
Tk−1Ak
)(I −R−1
k Ak)ν(4.4)
(see, e.g., [61, p. 48]). Note that for J = 1 (4.4) reduces to the two-grid iteration matrix:
E(J)TG = (I −R−1
J AJ)ν(I − PJ−1A
−1J−1P
TJ−1AJ
)(I −R−1
J AJ)ν . (4.5)
The convergence on finest level is governed by the spectral radius ρ(E(J)MG), or, in case
of two-grid, ρ(E(J)TG). In this chapter, we want to discuss assessment of these spectral radii
within the framework of a Fourier analysis, as may be developed for systems arising from
the discretization of scalar PDEs. It means that the eigenvectors of Ak are explicitly
known at each level and form the Fourier basis. We further assume that the smoother
shares the same set of eigenvectors; i.e., is also diagonal when expressed in this Fourier
basis.1 According to (4.2), N (ν)k will be diagonal as well for all ν.
Technically, a Fourier analysis is then possible if, expressing Pk−1 in both the coarse
(level k − 1) and fine (level k) Fourier basis, it has the form
Pk−1 =
p(k−1)1
p(k−1)2
. . .
p(k−1)lk−1
O
, (4.6)
where p(k−1)j are (possibly complex) vectors of size m(k)
j , j = 1, ..., lk−1. Note that this
form induces a block partitioning of Ak, Rk and Nk when these matrices are expressed
in Fourier basis. More precisely we write
Ak =
Λ(k,1)
Λ(k,2)
. . .
Λ(k,lk)
, Rk =
Γ(k,1)
Γ(k,2)
. . .
Γ(k,lk)
,
1Here we exclude cases for which the smoother is block diagonal as, e.g., when using red-black Gauss-Seidel relaxation for 5-pont discretizations of Poisson equation [61, Section 4.5].
60 Smoothing factor and actual multigrid convergence
N(2ν)k =
Σ(k,1)
Σ(k,1)
. . .
Σ(k,lk)
, (4.7)
where Λ(k,j) = diag(λ(k,j)1 , . . . , λ
(k,j)
m(k)j
), Γ(k,j) = diag(γ(k,j)1 , . . . , γ
(k,j)
m(k)j
) , and Σ(k,j) =
diag(σ(k,j)1 , . . . , σ
(k,j)
m(k)j
). Note that the block lk corresponds to all eigenmodes of Ak that
have no corresponding block in Pk−1; that is, all modes such that the associated eigenvec-
tor vk satisfies P Tk−1vk = 0. Whereas the separate treatment of the “non-prolongated”
block is not compulsory (it can, for instance, be merged with one of the regular blocks),
we adopt it here because such block (which can also be a group of “non-prolongated”
blocks put together) often arises in practice.
Observe that the partitioning induced by (4.6) associates in a same block other than
lk the different Fourier modes that, on the coarse-grid, are mapped by P Tk−1 onto the
same coarse Fourier mode. To develop our analysis, we don’t need to enter the details
about which modes are associated. It is important to note, however, that in usual setting
of Fourier analysis (see, e.g., [61,68]), inside each set of associated modes (that is, inside
each block other than lk), there is a unique mode classified as “low frequency”, all other
modes being labelled as “high frequency”. Moreover, the modes that belong to the block
lk are all classified as “high frequency”. Then, the smoothing factor is the worst factor
by which high frequency components are reduced per relaxation step; that is,
µ(k) = maxj=1,...,lk
maxi=1,..,m
(k)j
i is a “high frequency” mode
|1− γ(k,j)i
−1λ
(k,j)i | .
In our study, we assume that the ordering inside each block is such that
|1− γ(k,j)1
−1λ
(k,j)1 | ≥ |1− γ
(k,j)2
−1λ
(k,j)2 | ≥ · · · ≥ |1− γ
(k,j)
m(k)j
−1λ
(k,j)
m(k)j
| (4.8)
and report results with respect to
µ(k) = maxj=1,...,lk
maxi=2,..,m
(k)j if j<lk
i=1,..,m(k)lk
if j=lk
|1− γ(k,j)i
−1λ
(k,j)i |
= max(
maxj=1,...,lk−1
|1− γ(k,j)2
−1λ
(k,j)2 | , |1− γ
(k,lk)1
−1λ
(k,lk)1 |
). (4.9)
Clearly, µ(k) coincides with the classical smoothing factor if, inside each block j other
than lk, the low frequency component is also the one less efficiently relaxed by the
smoother. This corresponds to usual situations, but may be not true in whole generality.
Smoothing factor and actual multigrid convergence 61
Note, however, that one has µ(k) ≤ µ(k) as soon as each block other than lk contains
at least one low frequency mode. In the following, we call µ(k) the smoothing factor
without checking further if µ(k) = µ(k).
Eventually, observe that (4.3) implies 1 − σ(k,j)i λ
(k,j)i = (1 − γ
(k,j)i
−1λ
(k,j)i )2ν and
hence (4.8) is equivalent to
σ(k,j)1 λ
(k,j)1 ≤ σ
(k,j)2 λ
(k,j)2 ≤ · · · ≤ σ
(k,j)
m(k)j
λ(k,j)
m(k)j
.
4.3 V–cycle analysis and McCormick’s bound
We recall here the bound obtained in [38, Lemma 2.3, Theorem 3.4 and Section 5] (see
also [37], or [53] for an alternative proof). The equivalence of definition (4.11) with
(4.12) is proved in Theorem 2.6.
Theorem 4.1. Let E(J)MG and N
(ν)k , k = 1, . . . , J , be defined respectively by (4.4)
and (4.2) with A being symmetric positive definite and with Pk , k = 0, . . . , J − 1 ,
Ak , k = 0, . . . , J , and Rk , k = 1, . . . , J , satisfying the general assumptions stated in
Section 4.2.
Then, letting πAk = Pk−1A−1k−1P
Tk−1Ak, there holds
ρ(E(J)MG) ≤ 1− min
1≤k≤Jδ
(ν)k , (4.10)
where
δ(ν)k = min
vk∈Rnk
‖vk‖2Ak − ‖(I − N(ν)k Ak)vk‖2Ak
‖(I − πAk)vk‖2Ak(4.11)
= minvk∈Rnk
vTk N(2ν)k vk
vTk (A−1k − Pk−1A
−1k−1P
Tk−1)vk
. (4.12)
Moreover,
δ(ν)k
−1≤ 1ν
(δ
(1)k
−1+ ν − 1
). (4.13)
As already mentioned, it was shown in Chapter 2 that the McCormick’s bound is
the best bound for V-cycle multigrid among those characterized by a constant which is
a maximum over all levels of an expression involving only two consecutive levels at a
time. This latter feature is the key property that allows us, in the next section, to assess
the bound in standard Fourier analysis setting, and relate it to the smoothing factor.
62 Smoothing factor and actual multigrid convergence
4.4 Rigorous Fourier analysis for SPD problems
Let Ak = N(2ν)k
1/2Ak N
(2ν)k
1/2and Pk−1 = N
(2ν)k
−1/2Pk−1 , with corresponding block
structure
Ak =
Λ(k,1)
Λ(k,2)
. . .
Λ(k,lk)
, Pk−1 =
p(k−1)1
p(k−1)2
. . .
p(k−1)lk−1
O
,
where Λ(k,j) = diag(λ(k,j)i ) with λ
(k,j)i = σ
(k,j)i λ
(k,j)i . Setting
π(k,j) = p(k−1)j ( p(k−1)
j
HΛ(k,j)p(k−1)
j )−1 p(k−1)j
HΛ(k,j) ,
there holds
ρ(E(k)TG) = ρ
(((I − Pk−1A
−1k−1P
Tk−1Ak
)(I − N
(ν)k Ak)2
)= ρ
((I − Pk−1A
−1k−1P
Tk−1Ak
)(I − N
(2ν)k Ak)
)= ρ
((I − Pk−1A
−1k−1P
Tk−1Ak)(I − Ak)
)= max
(max
j=1,...,lk−1ρ(
(I − π(k,j))(I − Λ(k,j))), ρ(I − Λ(k,lk)
)),
and
δ(ν)k
−1= max
vk∈Rnk
vTk (A−1k − Pk−1A
−1k−1P
Tk−1)vk
vTk N(2ν)k
−1vk
= maxvk∈Rnk
vTk (I − Pk−1A−1k−1P
Tk−1Ak)A
−1k vk
vTk vk
= max(
maxj=1,...,lk−1
ρ(
(I − π(k,j)) Λ(k,j)−1), ρ(
Λ(k,lk)−1))
.
Now, for each individual block other than lk, the quantities one has to take the
maximum of may be assessed by applying the following theorem with Λ = Λ(k,j) and
p = p(k−1)j (this vector is not equal to zero since the block lk is not considered). Observe
that the assumption 0 < λi ≤ 1 is then not restrictive since I−Ak and (I−N (ν)k Ak)2 have
the same spectra, and hence the eigenvalues of Λ(k,j), being a subset of the eigenvalues
of Ak, belong to (0, 1] by virtue of our general assumptions.
Smoothing factor and actual multigrid convergence 63
Theorem 4.2. Let Λ = diag(λi) be a m ×m real matrix with 0 < λ1 ≤ λ2 ≤ · · · ≤λm ≤ 1 , and let p = (p1 . . . pm)T be a nonzero complex vector. Set
π = p (pHΛp)−1 pHΛ ,
ρTG = ρ(
(Im − π)(Im − Λ
) ),
and
δ−1 = ρ(
(Im − π) Λ−1).
Letting
α =m∑i=2
λ2i |pi|2
λ21‖p‖2
, (4.14)
thenλ2
1 + α(1− λ1/λ2)≤ δ ≤
(4α
)1/3
. (4.15)
Moreover, if |p1| > 0, letting
β =m∑i=2
λ2i |pi|2
λ21|p1|2
, (4.16)
then
λ1 +λ2 − λ1
1 + β≤ δ ≤ λ1 +
λm − λ1
1 + β(4.17)
and
λ1 +λ2 − λ1
1 + λ−12 λ1β
≤ 1− ρTG ≤ min
(λ1 +
λm − λ1
1 + λ−1m λ1β
, λ2
), (4.18)
whereas, if |p1| = 0,
δ = 1− ρTG = λ1 . (4.19)
Proof. Set λc = pHΛp =∑m
i=1 λi|pi|2. First, observe that, according to Lemma 2.2
in [41],
λ1|p1|2λ−1c λ−1
m +(1−λ1|p1|2λ−1c )λ−1
1 ≤ δ−1≤ λ1|p1|2λ−1c λ−1
2 +(1−λ1|p1|2λ−1c )λ−1
1 , (4.20)
and, similarly,
λ1|p1|2λ−1c ηm+ (1− λ1|p1|2λ−1
c )η1 ≤ ρTG ≤ λ1|p1|2λ−1c η2 + (1− λ1|p1|2λ−1
c )η1 , (4.21)
64 Smoothing factor and actual multigrid convergence
where ηi = (1 − λi). Equality (4.19) then readily follows. Moreover, (4.20) and (4.21)
can be further rewritten as
λ1|p1|2λ−1c
(λ−1m − λ−1
1
)+ λ−1
1 ≤ δ−1≤ λ1|p1|2λ−1c
(λ−1
2 − λ−11
)+ λ−1
1 , (4.22)
λ1|p1|2λ−1c (ηm − η1) + η1 ≤ ρTG ≤ λ1|p1|2λ−1
c (η2 − η1) + η1 . (4.23)
We now prove the inequalities (4.17) and (4.18). Note that
λ1|p1|2λ−1c =
λ1|p1|2∑mi=1 λi|pi|2
=
(1 + λ1
∑mi=2 λi|pi|2
λ21|p1|2
)−1
implies
(1 + ξ2)−1 ≤ λ1|p1|2λ−1c ≤ (1 + ξm)−1 , (4.24)
where ξi = βλ1/λi. Using these last inequalities in (4.22) and (4.23) one obtains (since
λ−11 ≥ · · · ≥ λ−1
m and η1 ≥ · · · ≥ ηm)
1ξm + 1
(λ−1m − λ−1
1
)+ λ−1
1 ≤ δ−1 ≤ 1ξ2 + 1
(λ−1
2 − λ−11
)+ λ−1
1 ,
1ξm + 1
(ηm − η1) + η1 ≤ ρTG ≤1
ξ2 + 1(η2 − η1) + η1 .
Hence, using ξi = β λ1/λi and ηi = 1− λi , i = 2,m , we have
1 + β
λm + λ1β≤ δ−1 ≤ 1 + β
λ2 + λ1β,
1− λ2m + λ2
1β
λm + λ1β≤ ρTG ≤ 1− λ2
2 + λ21β
λ2 + λ1β.
The inequalities (4.17) and (4.18) (except the second term in the minimum) readily
follow. To conclude the proof of (4.18), let ek be the k-th unit vector and let
v =
e2 if pHe2 = 0
e1 − e2
(λ1
1/2pHe1
λ21/2
pHe2
)otherwise .
Note that π Λ−1/2
v = 0 and Λ π = πH Λ . Hence,
ρTG = ρ(
(Im − π)2(Im − Λ
) )= ρ
((Im − π)
(Im − Λ
)(Im − π)
)= ρ
(Λ1/2 (Im − π)
(Im − Λ
)(Im − π) Λ−1/2
)= ρ
(Λ−1/2 (Im − π)H Λ1/2
(Im − Λ
)Λ1/2 (Im − π) Λ−1/2
)
Smoothing factor and actual multigrid convergence 65
≥vHΛ−1/2 (Im − π)H Λ1/2
(Im − Λ
)Λ1/2 (Im − π) Λ−1/2v
vHv
=vH(Im − Λ
)v
vHv
≥ 1− λ2 .
We next prove the left inequality (4.15). First observe that
λ1|p1|2λ−1c =
λ1|p1|2∑mi=1 λi|pi|2
= 1−∑m
i=2 λi|pi|2∑mi=1 λi|pi|2
≥ 1−∑m
i=2 λi|pi|2
λ1∑m
i=1 |pi|2≥ 1− λ1α
λ2
,
and hence, with (4.20), there holds
δ−1 ≤
(1− λ1α
λ2
)λ−1
2 +λ1α
λ2
λ−11 = λ−1
2
(1 + α(1− λ1
λ2
)
).
It remains to prove the right inequality (4.15). Note that, according to Theorem 3.2,
δ ≤ min(
1− ρTG , ‖π‖−2).
Hence, provided that
α ≤ 4(1− ρTG)2
‖π‖2 (4.25)
holds (we prove it below), we have
δ ≤ min(
1− ρTG ,1α
4(1− ρTG)2
)≤ max
x>0min
(x ,
1α
4x2
)≤(
4α
)1/3
.
We are thus left with the proof of (4.25), for which we use
‖π‖2 = ρ(p λ−1
c pHΛ2p λ−1c pH
)=‖p‖2 pHΛ2p
λ2c
= ‖p‖2∑m
i=1 λ2i |pi|2
λ2c
. (4.26)
According to (4.21), we have
ρTG ≥ 1− λ1 − λ1|p1|2λ−1c (λm − λ1) .
Hence, considering first the case where λ1 ≤ 1−ρTG2 and λ1 6= λm, there holds
λ1|p1|2λ−1c =
|p1|2λ1∑mi=1 λi|pi|2
≥ 1− λ1 − ρTGλm − λ1
≥ 1− ρTG2(λm − λ1)
≥ 1− ρTG2
, (4.27)
66 Smoothing factor and actual multigrid convergence
the last inequality following from 0 ≤ λ1 < λm ≤ 1. Note that (4.27) implies |p1|2 > 0.
The right inequality (4.25) follows then from
α =m∑i=2
λ2i |pi|2
λ21‖p‖2
≤ ‖p‖2∑m
i=1 λ2i |pi|2
(λ1|p1|2)2≤ 4
(1− ρTG)2‖p‖2
∑mi=1 λ
2i |pi|2
λ2c
,
together with (4.26). If λ1 = λm, we have
α =m∑i=2
|pi|2
‖p‖2≤ 1 ≤ 4
(1− ρTG)2‖π‖2 ,
the last inequality coming from the fact that π is a projector, and hence ‖π‖ ≥ 1. On
the other hand, when λ1 ≥ 1−ρTG2 one has (since λi ≤ 1)
α =m∑i=2
λ2i |pi|2
λ21‖p‖2
<4
(1− ρTG)2
m∑i=2
λ2i |pi|2
‖p‖2≤ 4
(1− ρTG)2≤ 4
(1− ρTG)2‖π‖2 ,
the last inequality coming from ‖π‖2 ≥ 1.
This theorem can be applied in the context of Fourier analysis, setting Λ = Λ(k,j)
and p = p(k−1)j , where Λ(k,j) and p(k−1)
j come from the block representation of Ak =
N(2ν)k
1/2Ak N
(2ν)k
1/2and Pk = N
(2ν)k
−1/2Pk. Hence, the main constants for block j < lk
at level k are
α(k,j)ν =
m(k)j∑i=2
λ(k,j)i
2|(p(k−1)
j )i|2
λ(k,j)1
2‖p(k−1)
j ‖2=
∑m(k)j
i=2 σ(k,j)i λ
(k,j)i
2|(p(k−1)
j )i|2(σ
(k,j)1 λ
(k,j)1
)2∑m(k)j
i=1 σ(k,j)i
−1|(p(k−1)
j )i|2(4.28)
and
β(k,j)ν =
m(k)j∑i=2
λ(k,j)i
2|(p(k−1)
j )i|2
λ(k,j)1
2|(p(k−1)
j )1|2=
m(k)j∑i=2
σ(k,j)i λ
(k,j)i
2|(p(k−1)
j )i|2
σ(k,j)1 λ
(k,j)1
2|(p(k−1)
j )1|2, (4.29)
where we use subscript ν to recall that these quantities inherit the dependence of σ(k,j)i
on the number of smoothing steps. Taking all blocks other than lk into account, we set
α(k)ν = max
j=0,...,lk−1α(k,j)ν and β(k)
ν = maxj=0,...,lk−1
β(k,j)ν . (4.30)
Considering the contribution of block lk to both 1 − δ(ν)k
−1, ρ(E(k)
MG) and ρ(E(k)TG), it is
given by 1− λ(k,lk)1 . It is not surprising that the contribution is the same since no coarse-
grid correction is performed on the corresponding modes, which therefore undergo only
the action of the smoother.
Smoothing factor and actual multigrid convergence 67
Our definition (4.9) of the smoothing factor entails
min(λ
(k,lk)1 , min
j=1,...,lk−1λ
(k,j)2
)= 1−
(µ(k)
)2ν. (4.31)
Hence, using successively the right inequality (4.18) (with second term in the minimum),
the results in [27, Section 7.2] ( for the proof of ρ(E(J)TG) ≤ ρ(E(J)
MG) ), the inequality
(4.10) and the left inequality (4.15) (with 1− λ1/λ2 bounded above by 1), one obtains
the following cascade of inequalities
(µ(J))2ν ≤ ρ(E(J)TG) ≤ ρ(E(J)
MG) ≤ 1 − min1≤k≤J
δ(ν)k ≤ max
1≤k≤J
(µ(k)
)2ν+ α
(k)ν
1 + α(k)ν
. (4.32)
Observe that if maxk=1,...,J µ(k) ≈ µ(J) (which often holds in practice) and if α(k)
ν is nicely
bounded at each level, these inequalities define a narrow interval containing both the
two-grid and V-cycle multigrid convergence factors. On the other hand, if α(k)ν is large
at some levels, the right inequality (4.32) becomes ineffective, and the right inequality
(4.15) further shows that 1 − min1≤k≤J δ(ν)k will be indeed close to 1. As observed in
Chapter 3, the actual convergence of the V-cycle may then scale poorly with the number
of levels.
Now, the smoothing factor µ(k) can be directly assessed from λ(k,j)i and γ
(k,j)i , i =
1, ...,m(k)j , j = 1, ..., lk. However, α(k)
ν and β(k)ν are related to the σ(k,j)
i , which are known
only via the relation
1− σ(k,j)i λ
(k,j)i = (1− γ
(k,j)i
−1λ
(k,j)i )2ν .
Hence, α(k)ν and β(k)
ν may be difficult to assess, and their dependence on ν is also unclear.
It is therefore not obvious to predict how the previous cascade of inequalities (4.32)
evolves with respect to this parameter. The easiest way to overcome this difficulty it to
use (4.10) combined with (4.13). The cascade of inequalities then becomes
(µ(J))2ν ≤ ρ(E(J)TG) ≤ ρ(E(J)
MG) ≤ 1− min1≤k≤J
δ(ν)k
≤ max1≤k≤J
δ(1)k
−1− 1
δ(1)k
−1− 1 + ν
≤ max1≤k≤J
(µ(k)
)2+ α
(k)1(
µ(k))2 + α
(k)1 + ν(1−
(µ(k)
)2). (4.33)
Note that the two rightmost bounds behave like O(ν−1), whereas the smoothing factor
alone suggests an exponential dependence via the left inequality (4.33). For the typical
example considered in Section 4.6, the actual behavior of both ρ(E(J)MG) and ρ(E(J)
TG)
is close to O(ν−1) (see Figure 4.1), indicating that the upper bounds provide a more
realistic estimate, at least in the considered case.
68 Smoothing factor and actual multigrid convergence
Now, in Theorem 4.3 below, we show that α(k)ν and β(k)
ν cannot increase with ν and
we further relate these constants to
α(k) = maxj=1,...,lk−1
∑m(k)j
i=2 γ(k,j)i
−1λ
(k,j)i
2|(p(k−1)
j )i|2(γ
(k,j)1
−1λ
(k,j)1
)2∑m(k)j
i=1 γ(k,j)i |(p(k−1)
j )i|2, (4.34)
β(k) = maxj=1,...,lk−1
m(k)j∑i=2
γ(k,j)i
−1λ
(k,j)i
2|(p(k−1)
j )i|2
γ(k,j)1
−1λ
(k,j)1
2|(p(k−1)
j )1|2. (4.35)
Note, however, that these expressions make sense only if, similarly to σ(k,j)1 λ
(k,j)1 =
min1≤s≤m(k)
j
σ(k,j)s λ
(k,j)s , one also has
γ(k,j)1
−1λ
(k,j)1 = min
1≤s≤m(k)j
γ(k,j)s
−1λ(k,j)s . (4.36)
This, in turn, holds if,
γ(k,j)i
−1λ
(k,j)i ≤ 2− min
1≤s≤m(k)j
γ(k,j)s
−1λ(k,j)s , i = 1, ...,m(k)
j . (4.37)
Indeed, (4.37) implies in particular min1≤i≤m(k)
j
γ(k,j)i
−1λ
(k,j)i ≤ 1, and further
|1− min1≤s≤m(k)
j
γ(k,j)s
−1λ(k,j)s | = 1− min
1≤s≤m(k)j
γ(k,j)s
−1λ(k,j)s
≥ max(1− γ(k,j)i
−1λ
(k,j)i , γ
(k,j)i
−1λ
(k,j)i − 1) ,
hence (4.36) by virtue of the ordering (4.8). Observe that (4.37) holds when the smoother
is scaled in such a way that the eigenvalues of R−1k Ak do not exceed 1. On the other
hand, if the smoother is related to some damping factor, the condition (4.37) is in fact
a constraint on the latter: assuming
γ(k,j)i
−1= ω γ
(k,j)i
−1,
it amounts to, taking all blocks other than lk into account,
ω ≤ min1≤j≤lk−1
2
max1≤s≤m(k)
j
γ(k,j)i
−1λ
(k,j)i + min
1≤s≤m(k)j
γ(k,j)i
−1λ
(k,j)i
.
Smoothing factor and actual multigrid convergence 69
Theorem 4.3. Let λi > 0 and γi > 0, i = 1, ...,m satisfy |1 − λ1γ−11 | ≥ |1 − λ2γ
−12 | ≥
· · · ≥ |1− λmγ−1m | and set
σ(ν)i = λ−1
i
(1− (1− λiγ−1
i )2ν)
, i = 1, ...,m ,
for some integer ν > 0 . Let
αν =∑m
i=2 σ(ν)i λ2
i |pi|2(σ
(ν)1 λ1
)2∑mi=1 σ
(ν)i
−1|pi|2
,
and, if |p1| > 0,
βν =m∑i=2
σ(ν)i λ2
i |pi|2
σ(ν)1 λ2
1|p1|2.
One has (1ν
)2
α1 ≤ αν ≤ α1 , (4.38)
and, if |p1| > 0,1νβ1 ≤ βν ≤ β1 . (4.39)
Moreover, if
max1≤i≤m
γ−1i λi ≤ 2− min
1≤i≤mγ−1i λi (4.40)
let
α =∑m
i=2 γ−1i λ2
i |pi|2(γ−1
1 λ1
)2∑mi=1 γi|pi|2
,
and, if |p1| > 0,
β =m∑i=2
γ−1i λ2
i |pi|2
γ−11 λ2
1|p1|2.
One has (2− ω
2ν
)2
α ≤ αν ≤ α , (4.41)
and, if |p1| > 0,2− ω
2νβ ≤ βν ≤ β . (4.42)
where ω = max1≤i≤m γ−1i λi.
Further, letting
φ = arccos
(|p1|√∑mi=1 |pi|2
),
one has, for any µ such that |1− γ−1i λi| ≤ µ, i = 2, ...,m
(1− µ)2 min2≤i≤m γimax1≤i≤m γi
(sinφγ−1
1 λ1
)2
≤ α ≤ ω2 max2≤i≤m γimin1≤i≤m γi
(sinφγ−1
1 λ1
)2
, (4.43)
70 Smoothing factor and actual multigrid convergence
and, if |p1| > 0,
(1− µ)2 min2≤i≤m γiγ1
(tanφγ−1
1 λ1
)2
≤ β ≤ ω2 max2≤i≤m γiγ1
(tanφγ−1
1 λ1
)2
. (4.44)
Proof. First, observe that we have
1− λiσ(1)i = (1− γ−1
i λi)2 , i = 1, ...,m ,
and, hence, the assumed ordering is equivalent to λ1σ(1)1 ≤ λ2σ
(1)2 ≤ · · · ≤ λmσ
(1)m .
Moreover, when (4.40) holds, one also has γ−11 λ1 = min1≤i≤m γ
−1i λi. We also note that
it is sufficient to prove inequalities (4.41) and (4.42) for ν = 1, the general case following
from (4.38) and (4.39), respectively.
Next, observe that
σ(ν)i
σ(1)i
=1− (1− γ−1
i λi)2ν
1− (1− γ−1i λi)2
=1− (1− s)ν
s=
ν∑k=0
(1− s)k (4.45)
with s = 1 − (1 − γ−1i λi)2 = λiσ
(1)i ∈ [0, 1], is a decreasing function of λiσ
(1)i . Hence,
since λ1σ(1)1 ≤ λ2σ
(1)2 ≤ · · · ≤ λmσ(1)
m , one has
σ(ν)i ≤
σ(ν)1 σ
(1)i
σ(1)1
, i = 1, ...,m .
The right inequalities (4.38) and (4.39) straightforwardly follow.
Similarly, sinceσ
(1)i
γ−1i
=1− (1− γ−1
i λi)2
γ−1i λi
= (2− γ−1i λi) , (4.46)
the equality γ−11 λ1 = min1≤i≤m γ
−1i λi implies
σ(1)i ≤
σ(1)1 γ−1
i
γ−11
, i = 1, ...,m ;
hence the right inequalities (4.41) and (4.42).
Next, fromσ
(ν)i
σ(1)i
=ν−1∑k=0
(1− γ−1i λi)2k
we conclude
1 ≤σ
(ν)i
σ(1)i
≤ ν ;
Smoothing factor and actual multigrid convergence 71
hence the left inequalities (4.38) and (4.39). Similarly, from (4.46) we have
2− ω ≤σ
(1)i
γ−1i
≤ 2
which in turn implies the left inequalities (4.41) and (4.42).
Finally, for the proof of (4.43) and (4.44) we first note that |1 − γ−1i λi| ≤ µ, i =
2, ...,m and the definition ω = max1≤i≤m γ−1i λi imply 1− µ ≤ γ−1
i λi ≤ ω. Hence
(1− µγ−1
1 λ1
)2 ∑mi=2 γi|pi|2∑mi=1 γi|pi|2
≤ α ≤(
ω
γ−11 λ1
)2 ∑mi=2 γi|pi|2∑mi=1 γi|pi|2
,
and, if |p1| > 0,
(1− µγ−1
1 λ1
)2 m∑i=2
γi|pi|2
γ1|p1|2≤ β ≤
(ω
γ−11 λ1
)2 m∑i=2
γi|pi|2
γ1|p1|2.
The conclusion follows since
sin2 φ =∑m
i=2 |pi|2∑mi=1 |pi|2
and tan2 φ =∑m
i=2 |pi|2
|p1|2.
This theorem shows that, from a qualitative viewpoint, it is sufficient to analyze
α(k,j) and β(k,j), which involve only γ(k,j)i , λ
(k,j)i and p(k−1)
j . Further, if, as often arises,
the smoother is well conditioned, all γ(k,j)i are approximately equal (they are all equal
for damped Jacobi smoothing). Then α(k,j)ν and β
(k,j)ν , j < lk, behave essentially like(
sinφ(k,j)
γ(k,j)1
−1λ(k,j)1
)2
and(
tanφ(k,j)
γ(k,j)1−1λ(k,j)1
)2
, respectively, where φ(k,j) is the angle between
the eigenvector associated to λ(k,j)1 and the range of the prolongation.
This allows to discuss the condition for having satisfactory two-grid convergence
and a satisfactory V-cycle convergence estimate via McCormick’s bound (4.10). Con-
sidering (4.17) and (4.18), only blocks for which λ(k,j)1 = σ
(k,j)1 λ
(k,j)1 is small have to
by analyzed carefully. This sounds logical: if all modes inside a block are efficiently
relaxed by the smoother, it does not matter that much how the restriction and prolon-
gation operate on these modes. Now, provided that the smoothing factor is bounded
away from 1, one will have nice two-grid convergence if and only if, for each block with
small γ(k,j)1
−1λ
(k,j)1 , the quantity λ
(k,j)1 β
(k,j)ν is reasonably bounded above; that is, if
tanφ(k,j) ≤ c ·(γ
(k,j)1
−1λ
(k,j)1
)1/2
. On the other hand, the condition is stronger for
having a nice V-cycle convergence estimate: this requires β(k,j)ν to be bounded above;
that is, tanφ(k,j) ≤ c · γ(k,j)1
−1λ
(k,j)1 .
Some heuristics present in the multigrid literature [18, p. 1573](see also [21, p. 4])
state: “Interpolation must be able to approximate an eigenvector with error bound
72 Smoothing factor and actual multigrid convergence
proportional to the size of the associated eigenvalue”. Our results give a more precise
interpretation of such statements. For mere two-grid convergence the tangent of the
angle between the eigenvector and the range of the prolongation should be proportional
to the square root of the eigenvalue, whereas guaranteed V-cycle convergence requires
it be proportional to the eigenvalue.
4.5 Semi-positive definite problems and local Fourier anal-
ysis
In this section we consider Fourier analysis for symmetric semi-positive definite linear
systems. Such extension is motivated by local Fourier analysis (also called local mode
analysis) that has a wider scope than (rigorous) Fourier analysis. The main idea is
the assessment of the two-grid convergence or of the smoothing factor without taking
boundary conditions into account. In practice, such approach is often equivalent to the
use of periodic boundary conditions and therefore leads to linear systems with non-trivial
null space.
Another way to interpret local Fourier analysis is to consider it as a limit case of
(rigorous) Fourier analysis for SPD problems on grids of increasing size (with, thus,
decreasing influence of boundary conditions on estimated parameters). Now, we pre-
viously observed for the SPD case that the angle between the range of prolongation
and an eigenvector of A should be proportional to the size of the eigenvalue. In the
limit case of local Fourier analysis, the modes belonging to the null space N (A) of A
should therefore be interpolated exactly, which also corresponds to a common practice.
It then follows that null space components seemingly play no role in the convergence,
and hence that Fourier analysis may be carried out ignoring these modes. This, in-
deed, is the common practice when assessing the two-grid convergence factor (see, for
instance, [61, p.109], [68, p.107] and the references therein).
In Theorem 4.4 below we give theoretical foundation to this approach with respect
to V-cycle multigrid, showing that McCormick’s bound on the convergence rate can also
be computed ignoring singular modes, or, more precisely, restricting the minimum in
(4.12) to vectors belonging to the range of Ak. Since (4.12) is at the root of the further
analysis developed in Theorems 4.2 and 4.3, the application of the results in Section 4.4
to local mode analysis is then straightforward.
Now, to state our theorem, we need to extend our definition of the V-cycle multigrid
algorithm in Section 4.2 to Ak possibly singular. The only potential difficulty comes
in fact with the bottom level matrix A0 whose inverse is needed. In Theorem 4.4, we
assume that instead one uses any matrix B0 such that A0B0A0 = A0. Such matrices are
called 1-inverse in [6], and one may check that if r0 ∈ R(A0), then A0B0r0 = r0. On
Smoothing factor and actual multigrid convergence 73
the other hand, to generalize (4.12), we need the inverse of the restriction of Ak to its
range. The most convenient way to express it is to use the Moore-Penrose inverse A+k
of Ak, since, if Ak = Xdiag(λi)XT , then A+k = Xdiag(λ+
i )XT with
λ+i =
λ−1i if λi 6= 0 ,
0 otherwise.
The expression of A+k is thus particularly simple when using the Fourier basis which
makes Ak diagonal.
Theorem 4.4. Let xn+1 = MG(b, A,xn, J) be the vector resulting from the application
of the multigrid algorithm with V-cycle at level J > 1, where A is symmetric semi-
positive definite, and, in case A0 is singular, where A−10 is exchanged for any matrix A(1)
0
such that A0A(1)0 A0 = A0. Assume that Pk , k = 0, . . . , J−1 , Ak , k = 0, . . . , J , and Rk ,
k = 1, . . . , J satisfy the general assumptions stated in Section 4.2 with ρ(I−R−1k Ak) < 1
being replaced by ρ(I −R−1k Ak) ≤ 1 , with (I −R−1
k Ak)z = λz for |λ| = 1 if and only if
z ∈ N (Ak) . Let PR(A),N (A) be the orthogonal projector onto the range of A.
If b ∈ R(A), then, for any solution x to (4.1),
PR(A),N (A) (x− xn+1) = E(J)MG (x− xn) = E
(J)MG PR(A),N (A) (x− xn)
for some matrix E(J)MG satisfying
ρ(E(J)MG) ≤ 1− min
1≤k≤Jδ
(ν)k ,
with
δ(ν)k = min
vk∈R(Ak)
vTk N(2ν)k vk
vTk (A+k − Pk−1A
+k−1P
Tk−1)vk
. (4.47)
Proof. Let qk = dim(N (Ak)). Observe that, Ak being non-negative definite,
Ak−1 = P Tk−1AkPk−1 is non-negative definite with qk−1 ≤ qk. Without loss of generality,
we can express all the matrices using bases of Rnk , k = J, ..., 0 such that, when qk > 0,
the first qk canonical vectors span N (Ak). Hence, Ak admits a block representation
Ak =
(Oqk,qk
ARRk
), (4.48)
with all but lower right blocks being empty if qk = 0.
74 Smoothing factor and actual multigrid convergence
Similarly, we partition
Rk =
(RNNk RNRk
RRNk RRRk
), N
(ν)k =
N (ν)k
NNN
(ν)k
NR
N(ν)k
RNN
(ν)k
RR
, Pk−1 =
(PNNk−1 PNRk−1
PRNk−1 PRRk−1
),
where all but lower right blocks of Rk, N(ν)k and Pk−1 become empty when qk = 0 , and
where PNNk−1 and PRNk−1 are empty when qk > 0 with qk−1 = 0. If qk > 0, there holds
Ak−1 = P Tk−1AkPk−1 =
(PRNk−1
TARRk PRNk−1 PRNk−1
TARRk PRRk−1
PRRk−1TARRk PRNk−1 PRRk−1
TARRk PRRk−1
).
Hence, in view of the form (4.48) and the fact thatARRk is SPD, one must have PRNk−1 = O.
It then follows that, for any nk × nk matrix
Bk−1 =
(∗ ∗∗ BRRk−1
),
one has
Pk−1Bk−1PTk−1 =
(∗ ∗∗ PRRk−1B
RRk PRRk−1
T
), (4.49)
This latter relation also holds for qk = 0, the blocks denoted by a star ∗ being then
empty.
Now, xn+1 = MG(b, Ak,xn, k) may be expressed as
xn+1 = xn +BJ(b−Axn), (4.50)
where the matrix BJ is defined from the recursion
B0 = A(1)0
Bk = N(2ν)k − (I − N
(ν)k Ak)Pk−1Bk−1P
Tk−1(I −Ak N
(ν)k ) , k = 1, . . . , J
(see, e.g., [65, Section 5.1] ; B−1k in this reference corresponds to Bk here).
Since there holds
I − N(2ν)k Ak =
Iqk ∗
I − N(2ν)k
RRARRk
(4.51)
with all but lower right blocks being empty if qk = 0, letting
BJ =
(∗ ∗∗ BRRJ
),
Smoothing factor and actual multigrid convergence 75
it follows from (4.49) that BRRJ may be computed from the recursion
BRRk = N(2ν)k
RR− (I − N
(ν)k
RRARRk )PRRk−1B
RRk−1 P
RRk−1
T(I −ARRk N
(ν)k
RR) ,
k = 1, . . . , J .
On the other hand, when A0 is singular, A0B0A0 = A0 holds for A0 of the form (4.48)
if and only if BRR0 = ARR0−1, whereas from (4.51) we deduce
I − N(2ν)k
RRARRk = (I − N
(ν)k
RRARRk )2 .
Hence Ek = I −BRRk ARRk obeys the recursion
E0 = O
Ek = (I − N(ν)k
RRARRk )
(I − PRRk−1(I − Ek−1) ARRk−1
−1PRRk−1
TARRk
)(I − N
(ν)k
RRARRk ) ,
k = 1, 2, . . . , J .
similar to (4.4); that is, corresponding to a multigrid scheme satisfying all assumptions
of Theorem 4.1, which therefore implies
ρ(EJ) ≤ 1− max1≤k≤J
δ(ν)k (4.52)
with
δ(ν)k = min
v∈Rnk−qk
vTNRRk v
vT (ARRk−1 − PRRk−1 A
RRk−1
−1PRRk−1
T )v.
Moreover, δ(ν)k = δ
(ν)k since
A+k =
(O
ARRk−1
)
with all but lower right block being empty when qk = 0.
Finally, using (4.50) and the fact that b ∈ R(A), there holds
x− xn+1 = (I −BJA)(x− xn),
and hence
PR(A),N (A) (x− xn+1) =
(O O
O I
)(I ∗
I −BRRk ARRk
)(x− xn)
76 Smoothing factor and actual multigrid convergence
=
(O O
O EJ
)(x− xn)
=
(O O
O EJ
)PR(A),N (A) (x− xn) ,
which, together with (4.52), concludes the proof.
4.6 Examples
4.6.1 Usual prolongations in 2D
In this subsection we show how the conclusions of Theorems 4.2 and 4.3 can be used
to analyze usual prolongation operators presented in [68]. More precisely, we assess the
parameter2
tan2 φ(j) =∑m
i=2 |(p(j))i|2
|(p(j))1|2, (4.53)
that characterizes the quality of a prolongation, according to (4.44). Indeed, see Sec-
tion 4.4, if tanφ(j) ≤ c ·(γ
(j)1
−1λ
(j)1
)1/2
, j = 1, ..., l, holds for not too large c, then
a smoothing factor bounded away from one guarantees optimal two-grid convergence,
whereas the condition tanφ(j) ≤ c · γ(j)1
−1λ
(j)1 leads to optimal V-cycle multigrid con-
vergence. Since γ(j)1 is often close to 1, it is thus critical to check the behavior of tanφ(j)
when λ(j)1 becomes close to 0.
Such a discussion presumes that the prolongation P (which fixes tanφ(j)) and the
coarse-grid matrix A (which determines λ(j)i ) are both known. Here, we develop a slightly
different approach and, for various prolongation operators (taken from [68]), we indicate
conditions that the eigenvalues of a potential system matrix A should satisfy, when used
with such prolongations, in order to lead to optimal two- and multigrid algorithms.
Instead of index j, we use a couple (θ1,θ2) of angles (for two-dimensional problems),
adopting the notation of [68]. Although in the context of (rigorous) Fourier analysis θ1
and θ2 can take only a finite number of values (depending on the mesh-size assumed),
we follow here the common practice and perform computations allowing all values inside
a fixed interval (0, π).
Before we characterize in Table 4.1 various prolongations taken from Table 6.2 in [68],
we illustrate with bilinear prolongation how the corresponding results can be derived.
From Table 6.2 in [68] we learn that the Fourier symbol for bilinear prolongation is
(1 + cos(θ1))(1 + cos(θ2)). This means that for a block characterized by a couple (θ1, θ2)2Here and in what follows we omit the grid number k, the discussion of this subsection does not
depend on the choice of a particular grid.
Smoothing factor and actual multigrid convergence 77
we have (with elements ordered not necessarily according to (4.8))
p(θ1,θ2) =
(1 + cos(θ1))(1 + cos(θ2))
(1 + cos(θ1 + π))(1 + cos(θ2))
(1 + cos(θ1))(1 + cos(θ2 + π))
(1 + cos(θ1 + π))(1 + cos(θ2 + π))
= 4
(1− sin2(θ1/2))(1− sin2(θ2/2))
sin2(θ1/2) sin2(θ2/2)
(1− sin2(θ1/2)) sin2(θ2/2)
sin2(θ1/2)(1− sin2(θ2/2))
. (4.54)
There, p(θ1,θ2) is one of the blocks (4.6) of P in the relevant Fourier basis. Now, small
values of tan2 φ(θ1,θ2) are possible when all but one entry of p(θ1,θ2) are small; that is,
when both θ1 and θ2 are close to 0 or π. Since vectors p(θ1,θ2), p(θ1+π,θ2), p(θ1,θ2+π) and
p(θ1+π,θ2+π) have same entries (ordered differently), in what follows we consider only
the situation when (θ1,θ2)→ (0, 0) (the same comment holds for other prolongations in
Table 4.1). Hence, for small θ1 and θ2,
p(θ1,θ2) =
O(1)
O(θ2
1
)O(θ2
2
)O(θ2
1θ22
)
,
which, together with (4.53), gives, for the ordering satisfying (4.8) (when all γi are
bounded below), tan2 φ(θ1,θ2) = O(θ2
1 + θ22
). The same data is given in Table 4.1 for
other prolongations coming from Table 6.2 in [68]. If we indicate O (θη1 + θη2) in the
third column it means that optimal two-grid convergence based on an optimal smooth-
ing factor occurs if and only if the eigenvalues of the matrix tend to zero only for θ1
and θ2 approaching 0 or π, and not faster than O(
(sinη θ1 + sinη θ2)1/2)
, whereas guar-
anteed optimal V-cycle convergence requires the eigenvalue going to 0 not faster than
O (sinη θ1 + sinη θ2).
For example, consider any of the usual discretizations of an isotropic laplace oper-
ator on a uniform grid. The eigenvalues λ(θ1,θ2) satisfy the already observed symmetry
λ(θ1,θ2) = λ(θ1+π,θ2) = λ(θ1,θ2+π) = λ(θ1+π,θ2+π). Moreover, for 0 ≤ θ1, θ2 ≤ π/2, the
eigenvalue λ(θ1,θ2) becomes small only when (θ1, θ2) is close to (0, 0), behaving in it
neighborhood as O(θ1
2 + θ22). In view of the results given in Table 4.1, and provided
that the smoother with bounded away from one smoothing factor is used, all considered
prolongations lead to an optimal two-grid cycle, and all but constant upwind prolonga-
tion are guaranteed optimal with V-cycle. Note that numerical experiments confirm the
78 Smoothing factor and actual multigrid convergence
prolongation Fourier symbol tan2 φ(θ1,θ2)
bilinear (1 + cos(θ1))(1 + cos(θ1)) O(θ21 + θ22
)bicubic (8 + 9 cos(θ1)− cos(3θ1)) O
(θ41 + θ42
)×(8 + 9 cos(θ2)− cos(3θ2))
biquintic (128− 150 cos(θ1) + 25 cos(3θ1)− 3 cos(5θ1)) O(θ61 + θ62
)×(128− 150 cos(θ2) + 25 cos(3θ2)− 3 cos(5θ2))
constant upwind (1 + exp(iθ1))(1 + exp(θ2)) O (θ1 + θ2)seven point 1 + cos(θ1) + cos(θ2) + cos(θ1 − θ2) O
(θ21 + θ22
)Table 4.1: Various prolongations coming from Table 6.2 in [68], with asymptotical
behavior of tan2 φ(j) for small values of θ1 and θ2.
suboptimal behavior of the V-cycle in this case [41].
4.6.2 2D Poisson
In this subsection we illustrate quantitative aspects of the cascades of inequalities (4.32)
and (4.33). We consider the linear system resulting from the bilinear finite element
discretization of the two-dimensional Poisson problem
−∆u = f in Ω = (0, 1)× (0, 1)
u = 0 in ∂Ω
on a uniform grid of mesh size h = 1/MJ in both directions. The matrix corresponds
then to the following nine point stencil−1 −1 −1
−1 8 −1
−1 −1 −1
. (4.55)
Up to some scaling factor, this is also the stencil obtained with 9-point finite difference
discretization. We assume MJ = 2JM0 for some integer M0 , allowing J steps of regular
geometric coarsening. We consider bilinear prolongation
Pk =
(Jk
Ink
)
where Jk corresponds to the natural interpolation associated with bilinear finite element
basis functions. The restriction P Tk corresponds then to “full weighting”, as defined in,
e.g. [61] 3. We consider damped Jacobi smoothing: Rk = ω−1Jacdiag(Ak) , with ωJac = 2/3;
that is, such that ω = maxλ∈σ(R−1k Ak) λ = 1. Since the stencil is preserved on all levels,
3up to some scaling factor; the scalings of the prolongation and restriction are unimportant whenusing coarse-grid matrices of the Galerkin type.
Smoothing factor and actual multigrid convergence 79
it is sufficient to consider only two successive grids; to alleviate notation, we therefore
let N = N(ν)k , A = Ak , P = Pk−1 , Ac = Ak−1 = P TAP and πA = πAk = PA−1
c P TA .
Now, we asses µ and αν using (rigorous) Fourier analysis. The eigenvectors of A are,
for m, l = 1, . . . ,M − 1 , the functions
u(M)m,l = sin(mπx) sin(lπy) (4.56)
evaluated at the grid points. The eigenvalue corresponding to u(M)m,l is
λ(M)m,l = 4(3sm + 3sl − 4smsl) (4.57)
where
sm = sin2(θ(m)) , sl = sin2(θ(l)) . (4.58)
with
θ(m) =mπ
2M.
The prolongation P satisfies (see, e.g., [61, p. 87])
P T
u
(M)m,l
u(M)M−m,M−l
−u(M)M−m,l
−u(M)m,M−l
= 4
(1− sm)(1− sl)
smsl
sm(1− sl)(1− sm)sl
u
(M/2)m,l
for 1 ≤ m, l ≤ M/2 − 1 , with P Tu(M)m,l = 0 for m = M/2 or l = M/2 . Hence, P has a
block structure (4.6) with blocks
pTm,l = 4(
(1− sm)(1− sl) smsl sm(1− sl) (1− sm)sl),
which also correspond to the bilinear prolongation (4.54) considered previously. A and
N (2ν) are block diagonal like in (4.7) with diagonal blocks given by, respectively,
Λm,l = diag(λ
(M)m,l , λ
(M)M−m,M−l , λ
(M)m,M−l , λ
(M)M−m,l
),
Σ(ν)m,l = 64 diag
1− (1− λ(M)c,s /12)ν
λ(M)c,s
(c,s)=(m,l),(M−m,M−l),(m,M−l),(M−m,l)
.
To assess αν via (4.41) and to evaluate the smoothing factor µ, we have to determine
the smallest eigenvalue of the block (m, l). We restrict ourself to l,m ≤M/2, for which
it is given by λm,l = Λm,l(1, 1). The results are extended to other couples (m, l), noting
that the blocks (l,m), (M − l,m), (l,M − m) and (M − l,M − m) lead to the same
set of eigenvalues for Λm,l with same corresponding entries of the prolongation vector
80 Smoothing factor and actual multigrid convergence
ν µ2ν ρ(E(J)TG) ρ(E(J)
MG) 1− δ(ν) δ(1)−1−1
δ(1)−1−1+ν
µ2+αµ2+α+ν(1−µ2)
1 0.25 0.25 0.271 0.333 0.333 0.6252 0.0625 0.083 0.121 0.2 0.2 0.455
Table 4.2: The estimates of different terms involved in inequalities (4.33) for ν = 1, 2.Two-grid and V-cycle convergence factors are assessed considering J = 7 and M0 = 2.
pm,l, and hence with same contributions to µ and α1. Transcribing the definitions (4.9),
(4.28) and (4.30) for l,m ≤M/2, we have
µ = max1≤m,l≤M/2
maxi=1,...,4 if l=M/2 or m=M/2
i=2,...,4 if l,m<M/2
∣∣∣∣1− Λm,l(i, i)12
∣∣∣∣ , (4.59)
αν = max1≤m,l≤M/2−1
∑4i=2(pm,l)2
iΛm,l(i, i)2Σ(ν)
m,l(i, i)
(Λm,l(1, 1)Σ(ν)m,l(1, 1))2
∑4i=1(pm,l)2
i (Σ(ν)m,l(i, i))
−1, (4.60)
(4.61)
From (4.59) one finds
µ ≤ 12.
On the other hand, Theorem 4.3 applies, yielding
αν ≤ α , (4.62)
with
α = maxm,l≤M/2−1
∑4i=2(pm,l)2
iΛm,l(i, i)2
Λm,l(1, 1)2∑4
i=1(pm,l)2i
, (4.63)
We show in Appendix A that
α ≤ 1 , (4.64)
whereas we show in Chapter 2 that, for this model problem,
δ(ν)−1 ≤ 1 +1
2ν, ν = 1, 2.
We are then able to report in Table 4.2 all quantities involved in inequalities (4.33),
using numerically computed values for ρ(E(J)TG) and ρ(E(J)
MG), and approximating αν with
its upper bound 1 (from (4.62), (4.64)); numerical investigation reveal that the latter is
relatively accurate since one has effectively α1 = 1 whereas, for ν > 1, αν is apparently
bounded below by 0.75. As a consequence, the rightmost upper bound (4.33) is in this
example sharper than the rightmost upper bound (4.32), and we display only the former.
Smoothing factor and actual multigrid convergence 81
1 2 3 4 5 6 7 8 9 1010
−7
10−6
10−5
10−4
10−3
10−2
10−1
100
101
ν
Figure 4.1: The dependence of ρ(E(J)MG) () and ρ(ETG) (×), as well as leftmost lower
bound (4.33) (∗) and rightmost upper bound (4.33) (+) , on the number ν of smoothingsteps.
The dependence of these quantities with respect to ν is further investigated on Fig-
ure 4.1. One sees that the O(ν−1) behavior of the upper bound (4.33) provides a more
realistic estimate than the lower bound µ2ν based on the smoothing factor only.
4.7 Conclusion
We have presented two cascades of inequalities (4.32) and (4.33) witch determine in-
tervals containing simultaneously the bound of McCormick, the V-cycle multigrid con-
vergence factor and the two-grid convergence factor on the finest level. The intervals’
limits depend on µ(k), which usually coincides with the smoothing factor on level k,
and on the additional quantity α(k)ν , which can be further bounded by a more simple
parameter α(k). This latter parameter depends essentially on the quotient of the sine
of the angle φ between prolongation and the corresponding smooth eigenvalue of Ak;
if it is small, the aforementioned intervals are narrow, indicating that the two-grid and
V-cycle convergence factors are both well reflected by the smoothing factor.
Assuming µ(k) reasonably small, we have further shown that the two-grid convergence
is optimal if and only if the cosine of φ has a bound proportional to the square root
of the corresponding eigenvalue of Ak, whereas McCormick bound predicts an optimal
V-cycle convergence if and only if cosine of φ has a bound proportional to the eigenvalue
itself. Using this observation, we have clarified the heuristics in the multigrid literature
which use such proportionality as a guideline in the design of multigrid solvers.
Finally, we have extended our analysis to positive semi-definite systems, as can arise
when using local Fourier analysis. In particular, if the system is compatible, we have
shown that the kernel modes of the problem can be ignored in the multilevel setting.
82 Smoothing factor and actual multigrid convergence
Appendix A
Here we sketch the proof of (4.64), or, equivalently, of
4∑i=2
(pm,l)2i Λm,l(i, i)2 − Λm,l(1, 1)2
4∑i=1
(pm,l)2i ≤ 0 , 1 ≤ l,m ≤M/2− 1 . (4.65)
Expressing (4.65) with respect to sl and sm one may check (using, for instance, computer
algebra tools) that
1256
(4∑i=2
(pm,l)2i Λm,l(i, i)2 − Λm,l(1, 1)2
4∑i=1
(pm,l)2i
)=−18slsm−172s2
l s2m+54(s2
l sm+sls2m)−9(s4
l +s4m)+194(s3
l s2m+s2
l s3m)−250s3
l s3m
−89(s4l s
2m+s2
l s4m)+88(s4
l s3m+s3
l s4m)−16s4
l s4m−54(s3
l sm+s3msl)+42(s4
l sm+s4msl)
=−s4l
((3−7sm)2+36s2
m(1−2sm)+4s2m(1−2sm)2
)−s4
m
((3−7sl)2+36s2
l (1−2sl)+ v4s2l (1−2sl)2
)−2slsm(9+86slsm−27(sl+sm)+27(s2
l +s2m)−97(s2
l sm+sls2m)+125s2
l s2m−8s3
l s3m)
and the proof is done if we show that, for 0 ≤ sm, sl ≤ 1/2, one has
g(sl, sm) = 9+86slsm−27(sl+sm)+27(s2l +s2
m)−97(s2l sm+sls2
m)+125s2l s
2m−8s3
l s3m ≥ 0 .
This inequality follows from
g(sl, sm) = f(sl, sm) +12slsm +
972slsm(1− 2sl)(1− 2sm) + 2s2
l s2m(1− 4slsm)
provided that
f(sl, sm) = (27− 72s2l )s
2m + (37sl − 27)sm + 9 + 27s2
l − 27sl ≥ 0 . (4.66)
Since the discriminant of this quadratic equation, namely
D(sl) = 7668s4l − 7668s3
l + 1009s2l + 918sl − 243 ,
is negative for 0 ≤ sl ≤ 1/2 , and since in the same equation the factor 27− 72s2l before
s2m is positive for the same interval, the inequality (4.66) and, hence, the result (4.65),
follow.
Chapter 5Algebraic analysis of aggregation-based multigrid
Summary
Convergence analysis of two-grid methods based on coarsening by (unsmoothed) aggre-
gation is presented. For diagonally dominant symmetric (M-)matrices, it is shown that
the analysis can be conducted locally; that is, the convergence factor can be bounded
above by computing separately for each aggregate a parameter which in some sense
measures its quality. The procedure is purely algebraic and can be used to control a
posteriori the quality of automatic coarsening algorithms. Assuming the aggregation
pattern sufficiently regular, it is further shown that the resulting bound is asymptoti-
cally sharp for a large class of elliptic boundary value problems, including problems with
variable and discontinuous coefficients. In particular, the analysis of typical examples
shows that the convergence rate is insensitive to discontinuities under some reasonable
assumptions on the aggregation scheme.
5.1 Introduction
We consider multigrid methods [61,27,67] for solving large sparse n× n linear systems
Ax = b (5.1)
with symmetric positive definite (SPD) system matrix A. Multigrid methods are based
on the recursive use of a two-grid scheme. A basic two-grid method combines the action
of a smoother, often a simple iterative method, and a coarse grid correction, which
corresponds to the solution of the residual equations on a coarser grid. The convergence
depends on the interplay between these two components and, when simple smoothers are
used, it relies essentially on the coarsening ; that is, on the way the fine grid equations
are approximated by the coarse system.
83
84 Algebraic analysis of aggregation-based multigrid
Here we consider coarsening by aggregation. In such schemes, the fine grid unknowns
are grouped into disjoint sets, and each set is associated with a unique coarse grid
unknown. Piecewise constant prolongation is then a common choice, which means that
the solution of the residual equation computed on the coarse grid is transferred back to
the fine grid by assigning the value of a given coarse variable to all fine grid variables
associated with it. This makes the coarse grid matrix easy to compute and usually as
sparse as the original fine grid matrix.
Aggregation schemes are not new and trace back to [11, 20]. They did not receive
much attention till recently because of the difficulty to obtain grid independent conver-
gence on their basis [59, p.p. 522–524], see also [69, p. 663], where an accurate three
grid analysis is presented for the model Poisson problem. This may be related to the
fact that piecewise constant prolongation does not correspond to an interpolation which
is at least first order accurate, as required by the standard multigrid theory [27, Sections
3.5 and 6.3.2].
That is why aggregation is often associated with smoothed aggregation, a procedure
in which a tentative piecewise constant prolongation operator is smoothed [63,64]. This
allows to develop an appropriate convergence theory, but, at the same time, some of
the attractive features of pure (unsmoothed) aggregation are lost, since the coarse grid
matrices are less sparse and more costly to compute.
In this chapter, we investigate such pure aggregation schemes based on piecewise
constant prolongation. They may indeed lead to two-grid methods with grid independent
convergence properties, as recently shown in [41] for model constant coefficient discrete
PDE problems. There is no contradiction with the above quoted results, whose focus is
on the convergence properties of two-grid methods used recursively in so-called V-cycle
scheme [61]. Indeed, aggregation based multigrid methods tend to scale poorly with
the number of levels when using simple V- or even W-cycles, even though the two-grid
scheme converges nicely [41, 48]. However, this may be cured using more sophisticated
K-cycles, in which Krylov subspace acceleration is used at each level [49]. It is also
possible to improve scalability by increasing the number of smoothing steps on coarser
levels [32].
Now, the (Fourier) analysis developed in [41] only addresses constant coefficient
problems with artificial (periodic) boundary conditions. Although there are numerical
evidences that aggregation based methods can be robust in presence of varying or dis-
continuous coefficients [48], this remains yet to be proved. On the other hand, it is also
lacking an analysis which would not only allow to assess a given aggregation scheme for
a problem at hand, but could also serve as a guideline in the development of aggregation
algorithms, in much the same way the coarsening strategies used in classical AMG meth-
ods may be derived from the objective to keep reasonably bounded some convergence
measure of the resulting two-grid scheme [15,53,58,59].
Algebraic analysis of aggregation-based multigrid 85
In this chapter, we fill these gaps by developing a convergence analysis which relates
the global convergence to “local” quantities associated with each aggregate. This analy-
sis is based on a general algebraic result which requires only the knowledge of a splitting
of the system matrix A satisfying some given properties, and we show how this splitting
can be constructed in a systematic way when the matrix is diagonally dominant. Fur-
ther, the needed local quantities are easy to compute solving an eigenvalue problem of
the size of the aggregate. They can also be assessed analytically in a number of cases.
This assessment reveals that the convergence is to a large extent insensitive to variations
or discontinuities in PDE coefficients if one can introduce some reasonable assumptions
on the aggregation scheme.
Moreover, as seen below, the bounds deduced in this way can often be shown asymp-
totically sharp provided that one assumes a simplified smoothing scheme with only one
damped Jacobi pre- or post-smoothing step. Hence, we do not only develop a qualita-
tive analysis, but also a quantitative one, complementary to Fourier analysis: this latter
allows to assess the benefit of more smoothing steps or increasing smoother quality, but
is restricted to constant coefficient problems on rectangular grids.
Returning to a qualitative viewpoint, it should be mentioned that, since the bound
depends only on local quantities, it is independent of the global properties of the under-
lying PDE such as (full) elliptic regularity. For instance, estimates derived in Section 5.4
do not need the assumption that the underlying domain is convex, and, in fact, allow
re-entering corners.
The presented results share some features with the analysis of element-based al-
gebraic multigrid (AMGe) approaches, as developed in [18, 21, 31, 33]. Convergence
estimates presented there are also local and can be used to guide the coarsening pro-
cess. The AMGe coarsening itself however differs substantially from aggregation. It
applies only to finite element problems and requires the knowledge of element matri-
ces, whereas the associated prolongation is denser than the basic piecewise constant
prolongation considered here.
The remainder of this chapter is organized as follows. The general framework of
aggregation-based two-grid methods is introduced in Section 5.2. The algebraic analysis
is developed in Section 5.3, and illustrated in Sections 5.4 and 5.5 on PDE problems
with, respectively, continuous and discontinuous coefficients. Concluding remarks are
given in Section 5.6.
5.2 Aggregation-based two-grid schemes
The coarsening procedure is based on the agglomeration of the unknowns of the system
(5.1) into nc non-empty disjoint sets called aggregates. The size of the k-th aggregate
is denoted by n(k) > 0. Note that some aggregation procedures (e.g., [48]) leave part
86 Algebraic analysis of aggregation-based multigrid
of the unknowns outside the coarsening process, for instance because the corresponding
row is strongly dominated by its diagonal element. As will be seen below, our analysis
gives theoretical support to this approach. Therefore, besides the nc regular aggregates
we define the (pseudo) 0-th aggregate as the (possibly empty) set of n(0) unknowns that
are left outside the coarsening process. For the ease of presentation, and without loss of
generality, we assume the ordering of the unknowns such that those belonging to (k+1)-
th aggregate have higher indices that those belonging to k-th aggregate, k = 0, ..., nc−1.
The regular aggregates are the variables of the next (coarse) level in the multigrid
hierarchy. Once they are defined, the n× nc prolongation matrix is given by
(P )ij =
1 if i belongs to j-th aggregate , j = 1, ..., nc0 otherwise .
(5.2)
Hence, setting 1m = (1 1 · · · 1)T , with m being the vector size, we have
P =
0 0 · · · 0
1n(1)
1n(2)
. . .
1n(nc)
. (5.3)
In what follows, we assume a slightly more general form of (5.3)
P =
0 0 · · · 0
p(1)
p(2)
. . .
p(nc)
(5.4)
with p(k) being a vector of size n(k). We shall see, however, that for the considered
examples the choice p(k) = 1n(k) is often the best (or even the only reasonable) choice.
Once the prolongation P is known, the nc×n restriction matrix is set to its transpose
and the nc × nc coarse grid matrix is given by the Galerkin formula Ac = P TAP . In
order to complete the definition of a two-grid scheme, one also needs to specify the pre-
and post-smoother matrices R1, R2, as well as the number ν1 and ν2 of pre- and post-
smoothing steps, respectively. The iteration matrix ETG of the two-grid cycle is then
given by
ETG = (I −R−12 A)ν2(I − P TA−1
c PA)(I −R−11 A)ν1 . (5.5)
The main objective of this chapter is the analysis of its spectral radius ρ (ETG) (that
is, its largest eigenvalue in modulus), which governs the convergence of the two-grid
Algebraic analysis of aggregation-based multigrid 87
scheme.
It is often convenient to define a “global” smoother X via the relation
I −X−1A = (I −R−11 A)ν1(I −R−1
2 A)ν2 . (5.6)
X has the same effect in one iteration as ν2 steps of post-smoothing followed by ν1 steps
of pre-smoothing. In what follows, we assume that X is SPD, which does not necessarily
requires the symmetry of R1 and R2.
5.3 Algebraic analysis
The starting point of our analysis is a notorious identity for the two-grid convergence
rate introduced in [23, Theorem 4.3]. We recall it up to a slight generalization in
Theorem 5.1 below. The generalization, that is based on the results in [44], allows for
nonsymmetric smoothing scheme, e.g., ν1 = 1 and ν2 = 0. It is somehow important
because the parameter µD for D = diag(A), which is investigated in the remainder of
this chapter, appears then directly connected to the convergence factor of a simplified
two-grid scheme with only 1 pre- or post-smoothing step.
Theorem 5.1. Let A be an n × n SPD matrix and let P be an n × nc matrix of rank
nc < n. Let R1, ν1 and R2, ν2 be such that X, defined by (5.6), is an n×n SPD matrix
and let ETG be the two-grid iteration matrix defined by (5.5).
Then, setting πX = P (P TXP )−1P TX, we have
ρ(ETG) = max(λmax(X−1A)− 1, 1− 1
µX
), (5.7)
where
µX = maxv∈Rn\0
vTX(I − πX)vvTAv
.
Moreover, for any n× n SPD matrix D, setting πD = P (P TDP )−1P TD and
µD = maxv∈Rn\0
vTD(I − πD)vvTAv
,
there holds
µX ≤(
maxv∈Rn\0
vTXvvTDv
)µD . (5.8)
In particular, if R1 = R2 = ω−1D with ω−1 ≥ λmax(D−1A), one has
ρ(ETG) = 1− 1µX
, (5.9)
88 Algebraic analysis of aggregation-based multigrid
with
µX ≤ ω−1 µD , (5.10)
where equality is reached when ν1 + ν2 = 1.
Proof. The equality (5.7) is a direct consequence of [44, Theorem 2.1 and Corol-
lary 2.1], combined with the assumptions that A and X are SPD, which implies
maxv∈Rn\0
vTX(I − πX)vvTAv
= λmax
(A−1/2X(I − πX)A−1/2
)= λmax
(A−1X(I − πX)
).
The inequality (5.8) follows from Corollary 2.2 in [44], setting in this latter Y = D,
LY = D1/2 and Q = πD.
To prove (5.9), observe that ω−1 ≥ λmax(D−1A) implies, together with (5.6),
λmax(X−1A) ≤ 1. Hence (5.7) gives (5.9) since it is known by [44, Theorem 2.1] that
µX ≥ 1.
The inequality (5.10) follows from (5.8) combined with
maxv∈Rn\0
vTXvvTDv
= ω−1 maxv∈Rn\0
vTωD−1vvTX−1v
= ω−1 maxv∈Rn\0
vTv − vT (I − ωA1/2D−1A1/2)vvTv − vT (I −A1/2X−1A1/2)v
≤ ω−1 ,
where the last inequality holds because I − A1/2X−1A1/2 = (I − ωA1/2D−1A1/2)ν1+ν2 .
Eventually, when ν1 +ν2 = 1, one has X = ω−1D, which implies X(I−πX) = ω−1D(I−πD), and, hence, that (5.10) is an equality.
When D is chosen independently of P , the first factor in the right hand side of (5.8)
depends only on the smoothing scheme. If R1 = RT2 = R and ν1 = ν2 = ν, setting
S =(I −R−1A
)ν , one has further
vTXvvTDv
≤ σ−1 ∀v ∈ Rn\0 ⇐⇒ ||Sv||2A ≤ ||v||2A − σ||v||2AD−1A ∀v ∈ Rn .
Hence, when D = diag(A) (the choice that is privileged in the rest of this work) this
quantity is nothing but the inverse of the smoothing factor in Ruge-Stuben analysis [59].
On the other hand, the second factor in the right hand side of (5.8) depends on P but
not on X, and keeping it bounded amounts to satisfying an approximation property.
Now, our analysis is based on the splitting of A as
A = Ab +Ar , (5.11)
Algebraic analysis of aggregation-based multigrid 89
where Ab and Ar are both symmetric nonnegative definite and Ab is block diagonal:
Ab =
A(0)
A(1)
. . .
A(nc)
, (5.12)
where A(k), k = 0, ..., nc, is of size n(k) × n(k).
As an example, consider a symmetric diagonally dominant matrix A with positive
diagonal entries (in particular, if all off-diagonal entries are nonpositive, the matrix is
an M -matrix). The matrices A(k), k = 0, ..., nc can be constructed by restricting the
matrix A to the unknowns belonging to the k-th aggregate and then by subtracting the
corresponding contribution C(k) = diag(ci) from its diagonal, in order to keep
Ar =
C(0) ∗ · · · ∗∗ C(1) · · · ∗...
.... . .
...
∗ ∗ · · · C(nc)
(5.13)
diagonally dominant, and, hence, nonnegative definite. Since A is diagonally dominant,
the contribution subtracted from the diagonal of each A(k) can be such that either each
row of Ab is weakly diagonally dominant; that is
(Ab)jj −n∑i=1i 6=j
|(Ab)ij | = 0, j = 1, ..., n ; (5.14)
or such that each row of Ar is weakly diagonally dominant; that is
(Ar)jj −n∑i=1i 6=j
|(Ar)ij | = 0, j = 1, ..., n ; (5.15)
or something in between; that is
(|(A)jj | −n∑i=1i 6=j
|(A)ij |) +n∑i=1i 6=j
|(Ab)ij | ≥ (Ab)jj ≥n∑i=1i 6=j
|(Ab)ij |, j = 1, ..., n . (5.16)
Once the splitting is known, the following theorem allows to estimate the “global”
approximation property constant µD by means of “local” quantities µ(k)D , k = 0, ..., nc.
Because each µ(k)D corresponds to a particular aggregate k, it may be seen as a measure
of this aggregate’s quality.
90 Algebraic analysis of aggregation-based multigrid
Theorem 5.2. Let A = Ab + Ar be an n × n SPD matrix, with Ab and Ar symmetric
nonnegative definite and Ab having the block-diagonal form (5.12). Let P be an n× ncmatrix of rank nc < n and of the form (5.4). Let
D =
D(0)
D(1)
. . .
D(nc)
(5.17)
be an n× n SPD matrix, set πD = P (P TDP )−1P TD and
µD = maxv∈Rn\0
vTD(I − πD)vvTAv
. (5.18)
Letting
µ(0)D =
0 if n(0) = 0
supv(0)∈Rn(0)\N (A(0))
v(0) T D(0)v(0)
v(0) T A(0)v(0)if n(0) > 0
and, for k = 1, ..., nc,
µ(k)D =
0 if n(k) = 1
supv(k)∈Rn(k)\N (A(k))
v(k) T D(k)(I − π(k)D )v(k)
v(k) T A(k)v(k)if n(k) > 1 ,
(5.19)
where
π(k)D = p(k)(p(k) T D(k)p(k))−1 p(k) T D(k) , (5.20)
there holds
µD ≤ maxk=0,...,nc
µ(k)D . (5.21)
Moreover, µ(0)D < ∞ if and only if n(0) = 0 or A(0) is SPD, and, for k = 1, ..., nc,
µ(k)D <∞ if and only if N (A(k)) ⊂ span
p(k)
, with, in the latter case,
µ(k)D =
0 if n(k) = 1
maxv(k)∈R(A(k))\0
v(k) T D(k)(I − π(k)D )v(k)
v(k) T A(k)v(k)if n(k) > 1 .
(5.22)
Proof. We first prove the if and only if result for k = 1, ..., nc, the case k = 0 being
trivial. The if statement assumes N (A(k)) ⊂ spanp(k)
which means that either
N (A(k)) = 0 or N (A(k)) = spanp(k)
. In the former case the supremum in (5.19)
becomes a maximum over Rn(k)\0 = R(A(k))\0, hence, (5.22) and µ(k)D <∞. In the
latter case, decomposing any vector that does not belong to N (A(k)) as v = αp(k) + w,
Algebraic analysis of aggregation-based multigrid 91
w ∈ R(A(k))\0, and using D(k)(I − π(k)D )p(k) = (D(k)(I − π(k)
D ))Tp(k) = 0, we have
µ(k)D = sup
v∈Rn(k)\N (A(k))
vTD(k)(I − π(k)D )v
vTA(k)v= max
w∈R(A(k))\0
wTD(k)(I − π(k)D )w
wTA(k)w,
leading to the same conclusions. The only if statement is proved assuming N (A(k)) *span
p(k)
and showing that µ
(k)D = ∞. Indeed, taking v = αu + w with w ∈
N (A(k))\spanp(k)
(exists by assumption) and u ∈ R(A(k)) leads to
µ(k)D = sup
α∈R\0
|wTD(k)(I − π(k)D )w + 2αuTD(k)(I − π(k)
D )w + α2uTD(k)(I − π(k)D )u|
α2uTA(k)u.
Since wTD(k)(I − π(k)D )w 6= 0 by construction of w, this last expression is unbounded
for α→ 0.
We now prove (5.21). Note that this inequality is obvious when µ(k)D = ∞ for at
least one k. Hence, without loss of generality we may assume µ(k)D finite for k = 0, ..., nc.
Moreover, since nc < n, there holds µD > 0.
Now, observe that
D(I − πD) =
D(0)
D(1)(I − π(1)D )
. . .
D(nc)(I − π(nc)D )
(5.23)
and, hence,
µD = maxv∈Rn\0
vTD(I − πD)vvTAv
= maxv∈Rn\0
vTD(I − πD)vvTAbv + vTArv
= maxv∈Rn\0
∑k=1,...,nc
v(k) T D(k)(I − π(k)D )v(k) + v(0) T D(0)v(0)∑
k=0,...,ncv(k) T A(k)v(k) + vTArv
. (5.24)
Let v∗ = (v(0)∗
Tv(1)∗
T· · · v(nc)
∗T
)T be the vector that realizes this maximum. Notice
that∑
k=0,...,ncv(k)∗
TA(k)v(k)
∗ > 0. Indeed, because of the boundness of µ(k)D , k =
0, ..., nc, the equality∑
k=0,...,ncv(k)∗
TA(k)v(k)
∗ = 0 would imply a zero numerator in the
right hand side of (5.24), whereas, since A is SPD, vT∗ Arv∗ > 0, which would further
lead to µD = 0. This latter contradicts our assumption nc < n.
Next, since µD is finite, as shown above, v(k)∗ ∈ N (A(k)) implies
v(k)∗
TD(k)(I − π(k)
D )v(k)∗ = 0 .
92 Algebraic analysis of aggregation-based multigrid
Therefore, since π(k)D = I when n(k) = 1 (entailing D(k)(I − π(k)
D ) = 0)
µD =
∑k=1,...,nc
v(k)∗
TD(k)(I − π(k)
D )v(k)∗ + v(0)
∗TD(0)v(0)
∗∑k=0,...,nc
v(k)∗
TA(k)v(k)
∗ + vT∗ Arv∗
≤∑
k=1,...,ncv(k)∗
TD(k)(I − π(k)
D )v(k)∗ + v(0)
∗TD(0)v(0)
∗∑k=0,...,nc
v(k)∗
TA(k)v(k)
∗
≤ maxk=0,...,nc
v(k)∗ /∈N (A(k))
v(k)∗
TD(k)(I − π(k)
D )v(k)∗
v(k)∗
TA(k)v(k)
∗
≤ maxk=0,...,nc
µ(k)D .
A practical consequence of this theorem is to show that nodes for which the corre-
sponding row is strongly dominated by its diagonal element may be kept outside the
aggregation process by putting them into the (pseudo) 0-th aggregate. The proposition
below presents a simple estimate of the pseudo aggregate’s quality based on diagonal
dominance excess of corresponding rows.
Proposition 5.1. Assume that A is diagonally dominant, that the splitting A = Ab+Arsatisfies (5.15) for j = 1, ..., nc and that D(0) = diag
(A)ii|i = 1, ..., n(0)
. If n(0) > 0,
one has
µ(0)D = max
v∈Rn(0)
vTD(0)vvTA(0)v
≤ maxi=1,...,nc
(A)ii2(A)ii −
∑nj=1 |(A)ij |
. (5.25)
Proof. Set ηi = 2(A)ii −∑n
j=1 |(A)ij | and note that if ηi = 0 at least for one
i ≤ nc, the inequality is trivially satisfied. Otherwise, observing that A(0) ≥ diag(ηi),
the inequality (5.25) follows.
Regarding aggregates 1, ..., nc , it is clear that the value of µ(k)D strongly depends on
p(k). In the theorem below we further indicate the scope of variation of the aggregate’s
quality if A(k) and D(k) are given, and determine the p(k) that leads to the best quality.
Theorem 5.3. Let A(k) and D(k) be, respectively, an n(k) × n(k) non-zero symmetric
nonnegative definite matrix and an n(k)× n(k) SPD matrix, with n(k) > 1. Let p(k) be a
non-zero vector of size n(k). Let
µ(k)D = sup
v∈Rn(k)\N (A(k))
vTD(k)(I − π(k)D )v
vTA(k)v, (5.26)
where π(k)D = p(pTD(k)p)−1pTD(k) and let λ1 ≤ λ2 ≤ · · · ≤ λn(k) be the eigenvalues of
D(k)−1A(k). Then,
λ−12 ≤ µ(k)
D ≤ λ−11 . (5.27)
Algebraic analysis of aggregation-based multigrid 93
Moreover, if
D(k)−1A(k)p = λ1 p , (5.28)
then
µ(k)D =
1λ2, (5.29)
and, assuming µ(k)D finite,
vTD(k)(I − π(k)D )v = µ
(k)D vTA(k)v v ∈ R(A(k))
if and only if
D(k)−1A(k)v = λ2v with vTD(k)p = 0 . (5.30)
Proof. Note that the case µ(k)D =∞ implies nonempty N (A(k)) and, hence, λ1 = 0.
The inequalities (5.27) are then trivially satisfied. Moreover, according to Theorem 5.2
we have then N (A(k)) * span p. Hence, if (5.28) holds, dim(N (A(k))) ≥ 2, which in
turn implies λ2 = 0 and, therefore, (5.29).
Now, consider µ(k)D < ∞ which, according to Theorem 5.2, implies N (A(k)) ⊂
span p, and, hence, λ2 > 0. If N (A(k)) is nonempty, then N (A(k)) = span p and
λ1 = 0, which in turn implies (5.28). Therefore, for all v ∈ R(A(k)), pTD(k)v = 0, and,
hence, π(k)D v = 0. Then, according to the Theorem 5.2, we further have
µ(k)D = max
v∈R(A(k))\0
vTD(k)(I − π(k)D )v
vTA(k)v= max
v∈R(A(k))\0
vTD(k)vvTA(k)v
= λ−12 .
In addition, a vector v reaches the maximum if and only if (5.30) holds.
Finally, we treat the case where N (A(k)) is empty and, hence, A(k) is invertible. Let
xi be the eigenvector of D(k)−1A(k) associated with the eigenvalue λi. To prove the left
inequality (5.27) we set
v =
x2 if pTD(k)x2 = 0
x1 −(
pTD(k)x1
pTD(k)x2
)x2 otherwise ,
and note that π(k)D v = 0. Injecting such v 6= 0 into (5.26) we find
µ(k)D ≥
vTD(k)vvTA(k)v
≥ λ−12 .
The right inequality (5.27) follows from
µ(k)D = max
v∈Rn(k)\0
vTD(k)(I − π(k)D )v
vTA(k)v≤ max
v∈Rn(k)\0
vTD(k)vvTA(k)v
= λ−11 .
94 Algebraic analysis of aggregation-based multigrid
Moreover, if p = x1, then xi, i = 1, ..., n(k) are also eigenvectors of A(k)−1D(k)(I−π(k)
D )
with corresponding eigenvalues λi such that λ1 = 0 and, for i > 1, λi = λ−1i . Since
µ(k)D is the smallest eigenvalue of A(k)−1
D(k)(I − π(k)D ), (5.29) follows. Moreover, (5.30)
holds if and only if v is an eigenvector A(k)−1D(k)(I−π(k)
D ) associated with λ−12 = µ
(k)D ,
which is in turn equivalent to vTD(k)(I − π(k)D )v = µ
(k)D vTA(k)v .
By way of illustration, consider a symmetric diagonally dominant M -matrix and
assume that the splitting A = Ab + Ar is based on the rule (5.14). Then, each A(k)
is singular with its null space equal to span1n(k). Theorem 5.2 then shows that one
has to use p(k) = 1n(k) to keep µ(k)D finite, in which case, by Theorem 5.3, µ(k)
D =
λ2(D(k)−1A(k))
−1. When the diagonal dominance is strict, the two side inequality
(5.16) indicates that there is some freedom in the choice of the diagonal entries of Ab,
and one may wonder how to exploit it at best. The following remarks give some clues
in this respect.
Remark 5.3.1 When A(k) is irreducible and diagonally dominant with nonpositive off-
diagonal entries, and when D(k) is a diagonal matrix, D(k)−1A(k) is an irreducible
M -matrix and, hence, an eigenvector whose components are all positive (e.g., 1n(k)) is
necessarily the eigenvector associated with the smallest eigenvalue, which is unique.
Remark 5.3.2 Consider a diagonally dominant M -matrix for which the splitting A =
Ab +Ar is based on (5.16). If the diagonal dominance is strict for some rows associated
with aggregate k, assuming p(k) = 1n(k) , a nice way to quickly obtain a useful estimate
consists in choosing diagonal entries as large as possible while satisfying (5.16) with the
additional constraint that 1n(k) is an eigenvector of D(k)−1A(k), so that the condition
ensuring (5.29) holds. In particular, when D(k) is a diagonal matrix, it amounts to using
A(k) = A(k)0 + ηD(k) with A
(k)0 satisfying (5.14) and with η being the largest constant
such that (5.16) still holds.
5.4 Discrete PDEs with constant and smoothly varying
coefficients
5.4.1 Preliminaries
We start considering matrices associated with the five point stencil−αy
−αx αd −αx−αy
with αx, αy > 0 and αd ≥ 2(αx + αy) (5.31)
on a rectangular grid of arbitrary shape. For such matrices we want to assess boxwise
aggregates with four nodes per aggregate (as on Figure 5.1(a)) and linewise aggregates
Algebraic analysis of aggregation-based multigrid 95
(a) (b) (c)
Figure 5.1: Examples of (a) boxwise, (b) linewise and (c) L-shaped aggregationpatterns.
with two, three and four nodes (as on Figure 5.1(b)). We select the splitting A = Ab+Arsatisfying (5.15). The prolongation vector is p(k) = 1n(k) , k = 1, ..., nc and, as can be
checked from (5.32) and (5.34) below, it is an eigenvector of D(k)−1A(k) associated with
the smallest eigenvalue δdα−1d , where δd = αd − 2(αx + αy) ≥ 0. Theorem 5.3 then
implies that µ(k)D = λ2(D(k)−1
A(k))−1 = αdλ2(A(k))−1.
Considering more specifically boxwise aggregates, we have
A(k) =
αx + αy −αx −αy 0
−αx αx + αy 0 −αy−αy 0 αx + αy −αx
0 −αy −αx αx + αy
+ δd I , (5.32)
and, hence,
µ(k)D =
2αx + 2αy + δd2 min(αx, αy) + δd
, (5.33)
whereas for linewise aggregation of size m in x direction
A(k) =
αx −αx
−αx 2αx. . .
. . . . . . −αx−αx αx
+ δd I (5.34)
and, hence, the following formula holds for m = 2, ..., 4 :
µ(k)D =
2αx + 2αy + δd
(2−√m− 2)αx + δd
. (5.35)
It follows that linewise aggregates of size 4 oriented in the direction of strong cou-
pling become more attractive than boxwise aggregates whenever max(αx, αy) > (2 +√
2) min(αx, αy) . Always choosing the best aggregate shape, we have then
µ(k)D ≤ 3 +
√2 . (5.36)
96 Algebraic analysis of aggregation-based multigrid
Since linewise aggregates of size 3 and 2 have better quality than linewise aggregates of
size 4, as can be concluded from (5.35), this upper bound holds for them as well.
5.4.2 Constant coefficients
We now discuss more specifically the five point finite difference approximation of
∂
∂x
(αx∂u
∂x
)+
∂
∂y
(αy∂u
∂y
)+ βu = f on Ω , (5.37)
with uniform mesh size h in both directions, where the boundary ∂Ω of the domain
Ω ∈ R2 is the union of segments parallel to the x or y axis and connecting the grid
nodes. Note that Ω is possibly not convex and may contain holes.
If the PDE coefficients αx, αy and β are constant, the above results allow to assess
aggregate’s quality for some typical aggregate shapes. It is also easy to extend the
reasoning to further aggregation schemes, leading to bound above (5.36) by a modest
constant if either coefficients are isotropic (αx = αy) or if one uses linewise aggregation
along the strong coupling direction. For instance, if αx = αy, (5.35) with m = 3 also
applies to L-shaped aggregates as illustrated on Figure 5.1(c).
Regarding Neumann boundary conditions, only the quality of aggregates that con-
tain boundary nodes is not covered by the above analysis. Again, however, isotropic
coefficients and linewise aggregates aligned with strong coupling yield bounds similar
to (5.33) and (5.35). For instance, if αx = αy and β = 0, boxwise aggregation near
a Neumann boundary result in matrices A(k) and D(k) that have the form analyzed in
Lemma 5.1 below, with α1 = α2 and α3 = α4 = 0 (boundary aligned with grid lines),
α2 = α3 = α4 = 0 (resorting corners), or α1 = α2 = α3 and α4 = 0 (re-entering corners).
As shown in this lemma, one has then µ(k)D ≤ 2 in the two former cases and µ
(k)D ≤ 2.23
in the latter, compared to µ(k)D = 2 away from the boundary.
Note that our analysis does not require all aggregates having the same shape, which
in fact seldom occurs with practical aggregation algorithms (see [48] for an example).
One should just take care that the global µD is not larger than desired because of a few
irregular aggregates, which in practice can be prevented by breaking them into smaller
pieces.
5.4.3 Smoothly varying coefficients
Consider now the same discrete PDE (5.37) but with smoothly varying coefficients.
Because the matrices A(k) and D(k) are local to the aggregate at hand, they are equal, up
to a O(h) perturbation, to the matrices A(k)0 and D(k)
0 corresponding to PDE coefficients
that are constant and equal to the mean value inside the aggregate. Furthermore, 1n(k)
remains the eigenvector of D(k)−1A(k) associated with the smallest eigenvalue either
Algebraic analysis of aggregation-based multigrid 97
αx = αy , δd = 0 αx = 10αy , δd = 0pairwise L-shaped boxwise linewise linewise boxwise
(size=3) (size=4)N µ
(k)D µD µ
(k)D µD µ
(k)D µD µ
(k)D µD µ
(k)D µD µ
(k)D µD
12 2 1.940 4 2.315 2 1.959 2.2 2.184 3.756 3.638 11 8.43124 2 1.984 4 2.377 2 1.989 2.2 2.196 3.756 3.744 11 10.18548 2 1.996 4 2.394 2 1.997 2.2 2.199 3.756 3.753 11 10.77896 2 1.999 4 2.399 2 1.999 2.2 2.200 3.756 3.755 11 10.943
Table 5.1: The value of µD and of its upper bound (5.21) for different grid sizes.
because β = 0 and, hence, N (A(k)) = span1n(k), or by using the trick suggested
at the end of Section 5.3 in Remark 5.3.2 (see also Remark 5.3.1). Hence, as shown
in Theorem 5.3, µ(k)D is the inverse of the second smallest eigenvalue of D(k)−1
A(k).
Since the eigenvalues of a matrix are continuous functions of its entries, it means that,
asymptotically (for h → 0), µ(k)D tends to the smallest eigenvalue of D(k)
0
−1A
(k)0 ; that
is, to the value obtained in the constant coefficient case. Therefore, the results of the
previous subsection carry over the variable coefficient case, at least when the mesh size
h is small enough.
5.4.4 Numerical example
We consider the linear system resulting from the five point finite difference discretization
of (5.37) on Ω = [0, 1]×[0, 1] with Dirichlet boundary conditions and constant coefficients
αx, αy and β = 0. The discretization is performed on a uniform rectangular grid of mesh
size h = (N + 1)−1 in both directions.
For the sake of simplicity, we let N be a multiple of 12, which allows that the whole
domain is covered with aggregates of the same shape. Using the rule (5.14), the matrices
A(k) and D(k) are the same for all aggregates. As a consequence, the quality estimate
µ(k)D is the same as well.
We consider first an isotropic situation (αx = αy). The columns from 2 to 7 of
Table 5.1 then give the values of µD and of its upper bound µ(k)D for three types of
aggregation pattern, presented on Figure 5.1. Observe that when mode are added to
an aggregate, its quality is not necessarily deteriorated, as can be seen comparing L-
shaped and box aggregates. We next consider in columns 8 to 13 an anisotropic situation
(αx = 10αy). One sees that boxwise aggregation is not recommended in this case.
5.4.5 Sharpness of the estimate
Numerical results in Table 5.1 indicate that the bound (5.21) on µD can be asymp-
totically sharp for N large enough. Moreover, as shown in Theorem 5.1, if only one
Jacobi smoothing iteration is performed, we further have ρ(ETG) = 1 − ωµ−1D . Hence,
98 Algebraic analysis of aggregation-based multigrid
a sharp estimate of µD further leads to a sharp estimate of the two-grid convergence
rate. The reader can wonder why and when this happens. This is what we investigate in
the present subsection, starting with the first question for the particular case of boxwise
aggregates.
Consider that the setting of the above example holds. Without loss of general-
ity, we assume in addition that αx ≥ αy. First, we recall that D(k)−1A(k)p(k) =
λ1(D(k)−1A(k)) p(k), and, hence, the vector vb = (1 1 − 1 − 1)T ∈ R(A(k)) that can
be checked to satisfy D(k)−1A(k)vb = λ2(D(k)−1
A(k)) vb reaches, according to Theo-
rem 5.3, the supremum in definition (5.19) of µ(k)D . Therefore, setting
v = (γ1 vb T γ2 vb T · · · γnc vb T )T ,
we locally reproduce the maximizing vectors for every aggregate. Moreover, setting
γ1 = γ2 = · · · = γN = −γN+1 = · · · = −γ2N = γ2N+1 = · · · = 1 we further make v take
the same value at every two connected nodes that belong to different aggregates. Hence,
since Ar have the form (5.13) with diagonal blocks being diagonal matrices, there holds
(Ar)ij ((v)i − (v)j) = 0 for all i and j. Therefore, setting σi =∑n
j=1(Ar)ij , and since
σi > 0 only for the unknowns near the boundary, there holds
vArv = −n∑
i,j=1
12
(Ar)ij ((v)i − (v)j)2 +n∑i=1
σi (v)i2 (5.38)
=n∑i=1
σi (v)i2
= 2N(αx + αy)
= 2N−1(αx + αy)α−1d vTDv . (5.39)
On the other hand, note that p(k) T D(k)vb = 0 implies πDv = 0, and, hence,
µD ≥vTD(I − πD)vvTAbv + vTArv
=vTDv
vTAbv + vTArv
=vTDv
µ(k)D
−1vTDv + vTArv
=µ
(k)D
1 + µ(k)D
vTArvvTDv
. (5.40)
It then follows from (5.39) that µD → µ(k)D for N →∞.
The following theorem is useful in extending this analysis to a more general frame-
work.
Algebraic analysis of aggregation-based multigrid 99
Theorem 5.4. Let A = Ab + Ar, P , D, µD and µ(k)D , k = 0, ..., nc, be defined as in
Theorem 5.2. Assume µ(k)D finite for k = 0, ..., nc and let, for n(0) > 0 and n(k) > 1
k = 1, ..., nc,
v0 ∈ arg maxv(0)∈Rn(0)\0
(v(0) T D(0)v(0)
v(0) T A(0)v(0)
),
vk ∈ arg maxv(k)∈R(A(k))\0
(v(k) T D(k)(I − π(k)
D )v(k)
v(k) T A(k)v(k)
), (5.41)
with vk = 1 otherwise. Let γk, k = 0, ..., nc, be real parameters, and set
v =(γ0θ−10 v0
T γ1θ−11 v1
T · · · γncθ−1nc vnc
T)T
, (5.42)
where θk =(vk
T A(k)vk)1/2
if n(k) > 1 and θk = 1 otherwise. Assume either that
vTArv ≤ εvTAbv , (5.43)
or that n(0) = 0, that A(k) is singular for k = 1, ..., nc and that
(c + v)TAr(c + v) ≤ ε(
maxk=1,...,nc
µ(k)D
)−1
vTDv (5.44)
for some vector c =(ξ1 p(1) T · · · ξnc p(nc) T
)T.
Then
µD ≥1
1 + ε
∑nck=0 γ
2kµ
(k)D∑nc
k=0 γ2k
. (5.45)
Proof. We first prove the lower bound (5.45) based on the assumption (5.43).
Starting with the equality (5.24) in the proof of Theorem 5.2 and setting v(k) = γkθ−1k vk
together with v = v , we have
µD ≥∑
k=1,...,ncv(k) T D(k)(I − π(k)
D )v(k) + v(0) T D(0)v(0)∑k=0,...,nc
v(k) T A(k)v(k) + vTArv(5.46)
≥ 11 + ε
∑k=1,...,nc
γ2kθ−2k vk
T D(k)(I − π(k)D )vk + γ2
0θ−20 v0
T D(0)v0∑k=0,...,nc
γ2kθ−2k vk
T A(k)vk
=1
1 + ε
∑k=1,...,nc
γ2kµ
(k)D∑
k=0,...,ncγ2k
,
where the last equality follows from θ−2k vk
T D(k)(I − π(k)D )vk = µ
(k)D .
100 Algebraic analysis of aggregation-based multigrid
Now, we prove the lower bound (5.45) based on the assumptions related to (5.44).
We may then assume that
(c + v)TAr(c + v) ≤ εvTAbv (5.47)
holds, this inequality being proved later. Since µ(k)D is finite, Theorem 5.2 implies
N (A(k)r ) ⊂ spanp(k), k = 1, ..., nc. From the singularity of A(k), k = 1, ..., nc, we
further conclude that N (A(k)r ) = spanp(k) and, hence,(
ξkp(k) + vk)T
A(k)(ξkp(k) + vk
)= vTkA
(k)vk . (5.48)
Moreover, using the definition (5.20) of π(k)D , we also have
(ξkp(k) + vk
)TD(k)(I − π(k)
D )(ξkp(k) + vk
)= vTkD
(k)(I − π(k)D )vk . (5.49)
Therefore, injecting v = c + v and v(k) = ξkp(k) + γkθ−1k vk into (5.46) and using (5.48)
and (5.49), the proof is finished as in the previous case.
We are thus left with the proof of (5.47). From Theorem 5.3 we conclude that
D(k)−1A(k)vk = λ2(D(k)−1
A(k))vk and p(k) T D(k)vk = 0. Therefore,
vkT D(k)(I − π(k)
D )vk = vTkD(k)vk ,
which implies vTkD(k)vk = µkD vk
T A(k)vk . Hence,
vTDv ≤(
maxk=1,...,nc
µ(k)D
)vT Abv ,
which, together with (5.44) implies (5.47).
Now, we return to the previous example and prove the asymptotical sharpness for
linewise aggregates of size m ≤ 4. As in the boxwise case, the vector (5.41) is the second
eigenvector of α−1d A(k) given by
vb =
(1√
2−1 1−√
2 −1)T if m = 4 ,
(1 0 −1)T if m = 3 ,
(1 −1)T if m = 2 .
Hence, choosing
v = (γ1 vb T γ2 vb T · · · γnc vb T )T ,
with γ1 = −γ2 = γ3 = · · · = γN/m = −γN/m+1 = γN/m+2 = · · · = 1, we further make v
take the same value at every two connected nodes that belong to different aggregates.
Algebraic analysis of aggregation-based multigrid 101
Next, using again (5.38) with first term in the right hand side vanishing, we have
vArv = 2Nm−1αy‖vb‖2 + αxN(‖(vb)1‖2 + ‖(vb)m‖2)
≤ 2N(αx + αy)‖vb‖2
= 2N−1(αx + αy)α−1d µ
(k)D vTAbv , (5.50)
and, hence, (5.45) holds with ε = N−1(αx + αy)α−1d µ
(k)D . The asymptotical sharpness
follows then from Theorem 5.4.
Considering a general situation, we note that a lower bound close to the upper bound
(5.21) can be proved via (5.45) if there exists a vector v of the form (5.42), such that
(a)∑nck=0 γ
2kµ
(k)D∑nc
k=0 γ2k
is close to maxk=0,...,nc µ(k)D ;
(b) ε, defined via (5.43) or (5.44), is small compared to 1 .
Now, the condition (a) can be satisfied by using large values of γ2k where µ(k)
D is large.
When all µ(k)D are the same, we trivially have
∑nck=0 γ
2kµ
(k)D∑nc
k=0 γ2k
= maxk=0,...,nc
µ(k)D ,
independently of the choice of γk. As illustrated in Section 5.5, the use of γ2k with
variable magnitude allows to prove asymptotical sharpness in the case where the µ(k)D s
are not all the same.
Condition (b) is more difficult to check. One may start from relation (5.38) and
look for a vector v of the form (5.42) such that (Ar)ij ((v)i − (v)j) = 0 for all i and
j. If such a vector exists, the first term in (5.38) is zero. Then, let Ωh = 1, ..., nc be
the set of all coarse unknowns and set ∂Ωh =i |σi =
∑nj=1 (Ar)ij 6= 0
. If as in the
previous example σi is positive only for unknowns near the boundary, then ∂Ωh is a set
of “boundary” unknowns. Assuming σi and (v)i, i = 1, ..., nc reasonably bounded, we
have
vTArv =n∑i=1
σi(v)2i = O(|∂Ωh|) ,
whereas, assuming γ2k , k = 1, ..., nc, bounded below,
vTAbv =nc∑k=0
γ2kθ−2k vTkA
(k)vk =nc∑k=0
γ2k = O(|Ωh|) ,
In a discretized PDE context, the ratio |∂Ωh|/|Ωh| usually becomes arbitrary small as
the mesh is refined.
102 Algebraic analysis of aggregation-based multigrid
Further, the lower bound (5.45) can be obtained using only a (given) set of aggregates
(numbered from 1 to nc for convenience), setting
v = (γ1θ−11 v1
T · · · γncθ−1nc vnc
T 0T · · · 0T )T . (5.51)
Then, (5.38) becomes
vArv = −∑i,j∈Ωh
12
(Ar)ij ((v)i − (v)j)2 +∑i∈Ωh
σi (v)i2 ,
where Ωh is the set of unknowns belonging to the first nc aggregates and σi =∑
j∈Ωh(Ar)ij .
Again, setting ∂Ωh = i | σi 6= 0 and repeating the steps described above, one obtains
µD ≥1
1 + ε
∑nck=0 γ
2kµ
(k)D∑nc
k=0 γ2k
,
with ε = O(|∂Ωh|/|Ωh|). In practice, it means that the upper bound (5.21) can also be
asymptotically sharp when the µ(k)D s are not all equal, providing that the aggregates for
which µ(k)D is maximal cover a significant part of the domain.
As an example, consider a scalar PDE discretized on a grid from which we can extract
a Ωh = N × N square of nodes with every node corresponding to the same stencil of the
form (5.31). Then, assuming that the whole square is covered with box aggregates as
on the Figure 5.1(a), the relations (5.33), (5.40) and (5.39) can be used (with N instead
of N) to show that
µD ≥1
1 + εµD (5.52)
with µD = 2αx+2αy+δD2 min(αx,αy)+δd
and ε = 2N−1(αx + αy)α−1d µD.
5.5 Discrete PDEs with discontinuous coefficients
5.5.1 Preliminaries
As in the previous section, our analysis is based on the aggregates’ quality, which in
turn involves the computation of the second smallest eigenvalue of small matrices. The
following lemma is helpful in this respect.
Lemma 5.1. Let
Ad =12
4α1 −2α1 −2α1
−2α1 3α1+α2 −α1−α2
−2α1 3α1+α3 −α1−α3
−α1−α2 −α1−α3 2α1+α2+α3
(5.53)
Algebraic analysis of aggregation-based multigrid 103
and
Dd = diag (4α1 2(α1+α2) 2(α1+α3) (α1+α2+α3+α4)) , (5.54)
where α1 > 0 and α2, α3, α4 > 0. Ad is positive semi-definite, and let λ2(D−1d Ad) be
the smallest nonzero eigenvalue of D−1d Ad.
Then,
λ2(D−1d Ad) ≥
5−√
178
, (5.55)
and, if α1 = α2 and α3 = α4, there holds
λ2(D−1d Ad) = min
(12,
3α1 + α3
4(α1 + α3)
). (5.56)
Moreover, if α1 ≥ α2, α3, α4, one has
λ2(D−1d Ad) ≥ β (5.57)
with β = λ2(D−1d Ad) (≈ 0.449) being evaluated for α1 = α2 = α4 = 1 and α3 = 0.
Furthermore,
λ2(D−1d Ad) ≥
12
if
α1 ≥ α2 = α3 = α4
or α1 = α2 ≥ α3 = α4
or α1 = α2 = α3 ≥ α4 .
(5.58)
Proof. We first prove (5.55). Since the diagonal entries of Dd are non-decreasing
functions of α4 and Ad does not depend on this latter, λ2(D−1d Ad) does not increase
with increasing α4. Hence, setting Cd = limα4→∞D−1d Ad, we have
λ2(D−1d Ad) ≥ lim
α4→∞λ2(D−1
d Ad) = λ2(Cd) , (5.59)
where
Cd =
(D−1r Ar ∗
0T 0
),
with
Ar =12
4α1 −2α1 −2α1
−2α1 3α1 + α2
−2α1 3α1 + α3
and Dr = 2
2α1
α1 + α2
α1 + α3
.
104 Algebraic analysis of aggregation-based multigrid
Hence, λ2(Cd) is the smallest eigenvalue of D−1r Ar . Now, assume without loss of gener-
ality that α3 ≥ α2 (one may see that they play a symmetric role in the definition of Adand Dd). Then, setting
Ar =12
4α1 −2α1 −2α1
−2α1 3α1 + α2
−2α1 3α1 + α2
and Dr = 2
2α1
α1 + α2
α1 + α2
,
we have,
λmin(D−1r Ar) = min
v∈R3\0
vTArvvTDrv
= minv∈R3\0
vT Arv + 12(α3 − α2) (v)3
2
vT Drv + 2(α3 − α2) (v)32
≥ min(λmin(D−1r Ar) ,
14
) . (5.60)
One may check that the set of eigenvalues of D−1r Ar is
3α1 + α2
4(α1 + α2),
38
+2α1 ±
√17α2
1 + 14α1α2 + α22
8(α1 + α2)
(e.g., by assessing the determinant of Ar − λDr for all λ belonging to the set1). Since
2α1 −√
17α21 + 14α1α2 + α2
2 ≥ (2−√
17)(α1 + α2) holds for α1, α2 > 0, the inequality
(5.55) follows.
On the other hand, if α1 = α2 and α3 = α4, the set of eigenvalues of D−1d Ad is given
by
0 , 12 ,
3α1+α34(α1+α3) ,
12 + 3α1+α3
4(α1+α3)
, which leads to (5.56).
To prove (5.57) we note that, as previously observed, λ2(D−1d Ad) does not increase
with increasing α4. Since α1 is the largest coefficient by assumption, setting α4 = α1
gives a worst case estimate. Next, we assume without loss of generality that α2 ≥ α3
(again, they play a symmetric role). Let then A0,0, A1,0 and A1,1 be the matrices defined
via (5.53) with α1 = α4 = 1 and the couple (α2, α3) given by, respectively, (0, 0), (1, 0)
and (1, 1); that is
A0,0 =
2 −1 −1
−1 32 −1
2
−1 32 −1
2
−12 −1
2 1
, A1,0 =
2 −1 −1
−1 2 −1
−1 32 −1
2
−1 −12
32
,
A1,1 =
2 −1 −1
−1 2 −1
−1 2 −1
−1 −1 2
.
1All eigenvalues explicitly given in this proof have been checked with computer algebra.
Algebraic analysis of aggregation-based multigrid 105
Similarly, let D0,0, D1,0 and D1,1 be the matrices defined via (5.54) with α1 = α4 = 1
and (α2, α3) being, respectively, (0, 0), (1, 0) and (1, 1); that is, D0,0 = diag(4 2 2 2) ,
D1,0 = diag(4 4 2 3) and D1,1 = diag(4 4 4 4) . Then,
Ad = (α1 − α2)A0,0 + (α2 − α3)A1,0 + α3A1,1
Dd = (α1 − α2)D0,0 + (α2 − α3)D1,0 + α3D1,1 .
Next, using the min-max theorem (e.g., [3, Lemma 3.13]), we have
λ2(D−1d Ad) = max
v∈R4\0minw⊥v
wTD−1/2d AdD
−1/2d w
wTw
= maxv∈R4\0
minz⊥v
zTAdzzTDdz
= maxv∈R4\0
minz⊥v
(α1 − α2)zT A0,0z + (α2 − α3)zT A1,0z + α3zT A1,1z
(α1 − α2)zT D0,0z + (α2 − α3)zT D1,0 + α3zT D1,1z
≥ maxv∈R4\0
min
(minz⊥v
zT A0,0z
zT D0,0z, min
z⊥v
zT A1,0z
zT D1,0z, min
z⊥v
zT A1,1z
zT D1,1z
)
Hence,
λ2(D−1d Ad) ≥ min
(min
z⊥D1,014
zT A0,0z
zT D0,0z, min
z⊥D1,014
zT A1,0z
zT D1,0z, min
z⊥D1,014
zT A1,1z
zT D1,1z
), (5.61)
where the second term in the minimum further becomes, since D1/21,0 14 belongs to
N (D−1/21,0 A1,0D
−1/21,0 ),
minz⊥D1,014
zT A1,0z
zT D1,0z= λ2
(D−1
1,0A1,0
)= β .
Therefore, the proof of (5.57) is done if we show that the second term in (5.61) is the
smallest. For this, we note that the vector z = (64 −34 33 −62)T is orthogonal to
D1,014 = (4 4 2 3)T and that zT A1,0z = 15861.5 with zT D1,0z = 34718. Hence, the
second term is smaller than 0.46. Further, the first term in (5.61) is larger than 0.46, as
can be concluded from positive definiteness of
A0,0 + D1,014(D1,014)T − 0.46D0,0 =
16.16 15 7 12
15 16.58 8 11.5
7 8 4.58 5.5
12 11.5 5.5 9.08
106 Algebraic analysis of aggregation-based multigrid
which implies zT A0,0z− 0.46zT D0,0z ≥ 0 for any z ⊥ D1,014. Similarly, the third term
in (5.61) is larger than 0.46 since
A1,1 + D1,014(D1,014)T − 0.46D1,1 =
16.16 15 7 12
15 16.16 8 11
7 8 4.16 5
12 11 5 9.16
is also positive definite. The positive definiteness can be proved, for instance, checking
that the determinants of upper left 1× 1, 2× 2, 3× 3 and 4× 4 blocks are positive.
Eventually, we prove (5.58). If α1 = α2 and α3 = α4, the inequality is already proved
in (5.56). If α2 = α3 = α4, taking Dd = diag(4α1 2(α1 +α2) 2(α1 +α2) 2(α1 +α2))
we have λk(D−1d Ad) ≥ λk(D−1
d Ad) ∈
0 , 12 ,
3α1+α24(α1+α2) ,
12 + 3α1+α2
4(α1+α2)
which leads to
the same conclusions. If α1 = α2 = α3, the set of eigenvalues of D−1d Ad is given by
0 , 12 ,
12 + 4α1±
√10α2
1+4α1α4+2α24
4(3α1+α4)
and, since α1 ≥ α4 implies 10α2
1 + 4α1α4 + 2α24 ≤
16α21, the inequality (5.58) follows.
5.5.2 Analysis
We consider the PDE (5.37) with piecewise constant isotropic coefficients (αx(x, y) =
αy(x, y)) and β = 0, and assume Dirichlet boundary conditions. As in the previous
section, we consider the five point finite difference approximation with uniform mesh size
h in both directions (mesh box integration scheme [42]), and assume that the boundary
∂Ω of Ω ⊂ R2 is the union of segments parallel to the x or y axis and connecting the grid
nodes. We aim at assessing boxwise aggregation as illustrated on Figure 5.1(a), which
was shown relevant for isotropic coefficients in the previous section.
Here we assume that the possible discontinuities match the grid lines. Hence, Ω is
a union of non overlapping subdomains Ωi in which the coefficients are constant, and
the boundary ∂Ωi of each Ωi is formed by segments aligned with grid lines and having
grid nodes as end points. To exclude some uncommon situations, we assume that every
two such end points are separated by a distance not less than 2h and that each box
aggregate contains at least one point which is interior to one subdomains. In practice,
this assumption is automatically met if the mesh size is small enough; in fact, it has to
be not larger than h0/2, where h0 is the size of the coarsest mesh that still correctly
reproduces the geometry of the problem.
The most general situation corresponding to this setting is then schematized on
Figure 5.2(a) where the central aggregate has one node interior to Ω1 and the opposite
node at the intersection of four subdomains: Ω1, Ω2, Ω3 and Ω4. With the splitting
satisfying (5.14), the corresponding aggregate’s matrices A(k) and D(k) are given by
Algebraic analysis of aggregation-based multigrid 107
(a) (b)
1 2
3 4
2
1
Figure 5.2: (a) general box aggregate situation with respect to discontinuities and(b) discontinuity nodes aggregated with white point nodes.
(5.53) and (5.54), respectively, with αi, i = 1, ..., 4, being the PDE coefficient in the
subdomain Ωi.
Because of the assumption (5.14) and of Theorem 5.3, aggregate’s quality µ(k)D is the
inverse of the second smallest eigenvalue of D(k)−1A(k). Lemma 5.1 then shows us the
following.
• The approach is robust in all cases, since, by (5.55), µ(k)D is always bounded above
independently of the relation between the coefficients αi.
• Nevertheless, from a practical viewpoint, (5.55) allows a significant decrease of ag-
gregate’s quality compared with the constant coefficient case. However, according
to (5.57), which implies µ(k)D ≤ 2.23 (compared with 2 in constant coefficient case),
a major deterioration is avoided when α1 ≥ α2, α3, α4. The latter condition is
satisfied if nodes belonging to several subdomains Ωi are always aggregated only
with nodes that belong to Ωi with largest PDE coefficient αi. Roughly speaking,
the rule may be summarized as “aggregate discontinuity nodes with those of the
strong coefficient region”.
• In many practical cases, no more than two subdomains are involved at a time
for a single aggregate, and either α1 = α2 = α3, or α1 = α2 and α3 = α4, or
α2 = α3 = α4 hold, as illustrated on Figure (5.2)(b). Then, if the rule above is
applied; that is, if α1 is in addition the largest coefficient, (5.58) applies and shows
that there is no deterioration at all compared with the constant coefficient case.
5.5.3 Numerical example
Consider the PDE (5.37) on a square domain Ω = [0, 1]× [0, 1] with β = 0,
αx(x, y) = αy(x, y) =
1 if x ≤ 1/2
d (> 1) if x > 1/2 .
108 Algebraic analysis of aggregation-based multigrid
(a) (b)
Figure 5.3: Two potential aggregation strategies for the numerical example.
and with Dirichlet boundary conditions. Consider the linear system (5.1) resulting from
its five point finite difference discretization (mesh box integration scheme [42]) on the
regular grid of mesh size h = N−1. Since discontinuities needs to be aligned with grid
lines, N has to be even. For simplicity of presentation, we further assume that it is a
multiple of 4. The number of unknowns being (N−1)×(N−1) (there is no unknown for
Dirichlet nodes), the grid cannot be covered with box aggregates only and the coarsening
is completed by pair and singleton aggregates. Then, the domain may be covered with
box aggregates starting from the left bottom corner (as on Figure 5.3(a)) or from the
right bottom corner (as on Figure 5.3(b)).
strategy (a) strategy (b)N max
k=0,...,ncµ
(k)D µD max
k=0,...,ncµ
(k)D µD
32 3.385 3.181 2 1.99364 3.385 3.286 2 1.998128 3.385 3.336 2 2.000256 3.385 3.361 2 2.000
Table 5.2: The value of µD and of its upper bound (5.21) for different aggregationstrategies and for d = 10.
Note that the quality of aggregates outside discontinuity is at most 2, as can be
concluded in the isotropic case (αx = αy) from (5.33) (for box aggregates) or from
(5.35) with m = 2 (for pair aggregates). The bound is therefore determined by the
quality of aggregates containing nodes on the discontinuity, which are given for d = 10
in Table 5.2. Observe that for the second strategy the convergence estimate is exactly
the same as in the constant coefficient case. For box aggregates, this follows from the
analysis in the previous subsection: the aggregates then obeys the “strong coefficient”
rule stated above. Regarding the first aggregation strategy, note that for box aggregates
one has
µ(k)D = λ2(D(k)−1
A(k))−1 =4(1 + d)
3 + d, (5.62)
Algebraic analysis of aggregation-based multigrid 109
using (5.56) with α1 = α2 = 1 and α3 = α4 = d. This is also true in the pairwise case,
since then
A(k) =
(1 −1
−1 1
), D(k) =
(4
2(d+ 1)
).
Note that (5.62) implies µ(k)D = 3.38 for d = 10 and µ
(k)D → 4 for d→∞.
5.5.4 Sharpness of the estimate
Table 5.2 indicates that, once again, the upper bound (5.21) is seemingly asymptotically
exact. In fact, the reasoning developed at the end of Section 5.4 shows that, asymptot-
ically, µD cannot be smaller than 2 for an isotropic (αx = αy) PDE (5.37) with β = 0
and a regular covering by box aggregates in at least one subdomain in which the PDE
coefficients are constant. Hence, our analysis is accurate when discontinuity nodes are
aggregated with nodes in strong coefficient region, since then µ(k)D ≤ 2.23. If, in addition,
µ(k)D ≤ 2, like in the numerical example above, then the bound is asymptotically sharp.
It is more challenging to show the sharpness when µ(k)D is significantly larger than
2 for some aggregates along discontinuity, essentially because the proportion of such
aggregates is O(h) or less. Nevertheless, it is interesting to confirm that, as seen in
Table 5.2, such a limited amount of low quality aggregates is sufficient to affect the
global convergence, and hence that the rule “aggregate discontinuity nodes with those
of the strong coefficient region” has some practical relevance.
In this view, we prove the sharpness of our estimate for the numerical example above
with the first aggregation strategy (depicted on Figure 5.3(a)), which does not follow
the “strong coefficient” rule. Note that, using the same trick as explained at the end of
Section 5.4, a similar lower bound on µD can be obtained in more complicated examples
whose domain would contain a rectangular region with two subdomains separated by a
line in the middle and covered similarly with box aggregates.
To apply Theorem 5.4, we need to construct two vectors v and c such that
∑nck=0 γ
2kµ
(k)D∑nc
k=0 γ2k
→ maxk=1,...,nc
µ(k)D for N →∞ , (5.63)
whereas ε, defined by (5.44), goes to 0 as N becomes large. In the example under
investigation, there are some pair and singleton aggregates (see Figure 5.3), but we limit
the support of both vectors to the (2`+1)×(2`+1) box aggregates, where ` = N/4−1. We
identify each such aggregate k with a couple (i(k)x , i
(k)y ) of indices, 1 ≤ i(k)
x , i(k)y ≤ 2`+ 1,
such that (i(k)x +1, i(k)
y ), (i(k)x −1, i(k)
y ), (i(k)x , i
(k)y +1) and (i(k)
x , i(k)y −1) are, respectively, its
right, left, top and bottom neighboring aggregates. Note that the center of the domain
is a node belonging to aggregate (`+ 1, `+ 1) and that discontinuity aggregates satisfy
i(k)x = `+ 1.
110 Algebraic analysis of aggregation-based multigrid
Since p(k) = 1n(k) , the vector vk from Theorem 5.4 is given by the eigenvector
of D(k)−1A(k), associated with the second smallest eigenvalue λ2(D(k)−1
A(k)); that
is, by vd = (τ 1 τ 1)T for discontinuity aggregate, with τ = −(d + 1)/2, and by
vo = (−1 1 − 1 1)T for the ordinary ones. The corresponding local energy (semi-
) norms are given by θ2d = vTdA
(k)vd = (3 + d)2/2 for discontinuity aggregates, by
θ2o = vTo A
(k)vo = 8 for the aggregates on the left of the discontinuity line and by d θ2o
for those on the right of it.
Then, the vector v is defined by ( v(1) T , v(2) T , · · · v(nc) T )T with
v(k) = `−1(`− |`+ 1− i(k)y |)×
τ2 `−1vo if 1 ≤ i(k)
x < `+ 1
vd if i(k)x = `+ 1
12`−1vo if `+ 1 < i
(k)x ≤ 2`+ 1
0 otherwise ,
(5.64)
and the vector c corresponds to ( c(1) T , c(2) T , · · · c(nc) T )T with
c(k) = (`− |`+ 1− i(k)x |+
12
)(`− |`+ 1− i(k)y |)`−2 ×
τ14 if 1 ≤ i(k)
x < `+ 1
0 if i(k)x = `+ 1
14 if `+ 1 < i(k)x ≤ 2`+ 1
0 otherwise .
From (5.64) we conclude that
γ2k = `−2(`− |`+ 1− i(k)
y |)2 ×
τ2
4 `−2 θ2
o if 1 ≤ i(k)x < `+ 1
θ2d if i(k)
x = `+ 114`−2 dθ2
o if `+ 1 < i(k)x ≤ 2`+ 1
0 otherwise ,
(5.65)
and, setting
s(`) =2`−1∑i=1
(`− |`− i|)2 =∑i=1
(i2 + (i− 1)2
)= `(`+ 1)(2`+ 1)/3− `2 ,
there holds
∑k: i
(k)x =`+1
γ2k = θ2
d
∑1<i
(k)y <2`+1
`−2(`− |`+ 1− i(k)y |)2 = θ2
d`−2s(`) ,
∑k: i
(k)x 6=`+1
γ2k = θ2
o
d+ τ2
4
∑1<i
(k)y <2`+1
1≤i(k)x <`+1
`−4(`− |`+ 1− i(k)y |)2 = θ2
o
d+ τ2
4`−3s(`) .
Algebraic analysis of aggregation-based multigrid 111
Hence,∑
k γ2kµ
(k)D = (1 +O(`−1))
∑k: i
(k)x =`+1
γ2kµ
(k)D , entailing (5.63) since µ(k)
D is maxi-
mal for i(k)x = `+ 1.
On the other hand, observe that c + v takes the same value at any two connected
nodes belonging to aggregates (i(k)x , i
(k)y ) and (i(k)
x + 1, i(k)y ). Moreover, c+ v vanishes on
the boundary of the region delimited by box aggregates. Hence, the only contribution
to (c + v)TAr(c + v) as expressed by (5.38) comes from connections between (i(k)x , i
(k)y )
and (i(k)x , i
(k)y + 1). In this latter case, let j1 and j2 be two connected nodes belonging
to aggregates (i(k)x , i
(k)y ) and (i(k)
x , i(k)y + 1), respectively, with i
(k)y ≤ 2`. For every box
aggregate k, let k+ (resp. k−) be the set of two nodes belonging to this aggregate with
larger (resp. smaller) abscise. One then has
|(c+v)j1−(c+v)j2 | =
τ`−2(`− |`+ 1− i(k)x |) if 1 ≤ i(k)
x < `+ 1 and j1 ∈ k−τ`−2(`− |`+ 1− i(k)
x |+ 1) if 1 ≤ i(k)x < `+ 1 and j1 ∈ k+
τ`−1 if i(k)x = `+ 1 and j1 ∈ k−
`−1 if i(k)x = `+ 1 and j1 ∈ k+
`−2(`− |`+ 1− i(k)x |+ 1) if `+ 1 < i
(k)x ≤ 2`+ 1 and j1 ∈ k−
`−2(`− |`+ 1− i(k)x |) if `+ 1 < i
(k)x ≤ 2`+ 1 and j1 ∈ k+ .
Therefore, using (5.38) with, this time, the first term being nonzero and the second one
vanishing because of the limited scope of c + v, we have
(c + v)TAr(c + v) =(τ2 +
1 + d
2
) ∑i(k)x =`+1
1≤i(k)y ≤2`
`−2
+(τ2 + d
) ∑1≤i(k)x <`+1
1≤i(k)y ≤2`
`−4(
(`− |`+1−i(k)x |+1)2 + (`−|`+1−i(k)
x |)2)
=(2τ2 + 1 + d
)`−1 + 2
(τ2 + d
)`−3
∑1≤i(k)x <`+1
(i(k)x
2+(i(k)x − 1
)2)
=(2τ2 + 1 + d
)`−1 + 2
(τ2 + d
)`−3s(`) ,
whereas
vTDv =vTd vd(2d+ 2)∑
1<i(k)y <2`+1
i(k)x =`+1
`−2(`− |`+ 1− i(k)y |)2
+ vTo vo(d+ τ2)∑
1<i(k)y <2`+1
1≤i(k)x <`+1
`−4(`− |`+ 1− i(k)y |)2
=4((τ2 + 1)(d+ 1) + (d+ τ2)`−1
)`−2s(`) .
112 Algebraic analysis of aggregation-based multigrid
Hence, vTDv = O(`) whereas (c + v)TArest(c + v) = O(1), showing with (5.55) that
(5.44) holds with ε = O(`−1) , and therefore, together with (5.63), proving the asymp-
totical sharpness of the estimate.
5.6 Conclusion
We have developed an analysis of an aggregation-based two-grid method for SPD linear
systems. When the system matrix is diagonally dominant, an upper bound on the
convergence factor can be obtained in a purely algebraic way, assessing locally and
independently the quality of each aggregate by solving an eigenvalue problem of the size
of the aggregate. Our analysis also shows that nodes for which the corresponding row
is strongly dominated by its diagonal element can be safely kept outside the coarsening
process (see Proposition 5.1).
We have applied our bound to scalar elliptic PDE problems in two dimensions,
showing that aggregation-based two-grid methods are robust if
• in the presence of anisotropy, one uses linewise aggregates aligned with the direc-
tion of strong coupling;
• in the presence of discontinuities, one avoids mixing inside an aggregate nodes
belonging to a strong coefficient region or its boundary with nodes interior to a
weak coefficient region.
Furthermore, we have shown that the bound is asymptotically sharp when a significant
part of the domain is regularly covered by box or line aggregates of the same shape.
Note that we have conducted the analysis in two dimensions for the sake of simplicity.
The same type of analysis can be developed for three dimensional problems, leading to
similar conclusions.
Our results may also have an impact on practical aggregation schemes. Because of
the above mentioned sharpness, it is indeed sensible to expect that aggregation methods
can be improved by improving aggregates’ quality. And because aggregates’ quality is
cheap to assess, this parameter can effectively be taken into account in the design of
aggregation algorithms. For instance, one may a posteriori check aggregates’ quality
and break low quality aggregates into smaller pieces. It is also possible, in a greedy-
like approach, to decide whether a node (or a group of nodes) should be added to an
aggregate according its impact on the aggregate’s quality and/or select the neighboring
(sets of) nodes that are the most favorable in this respect. These practical aspects are
subject to further research.
Chapter 6Fourier Analysis of aggregation-based two-grid
method for edge element
Summary
We consider Reitzinger and Schoberl multigrid method for curl-curl problems discretized
with edge finite elements. We perform a Fourier analysis of its two-grid variant and show
that the corresponding convergence rate can be bounded independently of the problem
size. This result is also compared with the actual two-grid convergence, indicating that
the analysis is accurate. Some numerical experiments are further performed in multigrid
setting with various cycling strategies, showing that an optimal implementation of the
method may be obtained when using the K-cycle.
6.1 Introduction
We consider multigrid methods for linear systems resulting from the discretization with
edge elements of
curl(α curl(E)) + βE = f on Ω , (6.1)
where α, β > 0 and Ω ⊂ R3. This problem arises when the vector potential is computed
in magnetostatics, when time-harmonic formulation of Maxwell’s equations is used, or
when eddy current approximation is considered (see [8, 5] and the references therein).
Note that, since edge element discretization is performed on (6.1), the degrees of freedom
in the linear system are associated with edges.
It is well known that standard multigrid techniques, if applied to such discretized
problems, have poor convergence properties. When the multigrid hierarchy is induced
by the refinement of an underlying coarse mesh, as in geometric multigrid, it is further
proved in [77] that a two-grid method can not be optimal if based on a simple point
smoother (like standard Jacobi or Gauss-Seidel). Modifications of standard multigrid by
either using special smoothing techniques [29, 2] or by decoupling multilevel hierarchies
113
114 Fourier Analysis of aggregation-based two-grid method for edge element
for edge and node unknowns [30, 35, 4] have been proposed recently to overcome this
difficulty.
Here we consider more precisely an algebraic multigrid method based on the coarsen-
ing by aggregation of edge unknowns, as introduced by Reitzinger and Schoberl in [52].
The main idea behind this approach is to perform the aggregation of edge unknowns
so that it also corresponds (via a given edge-node incidence matrix) to an aggregation
of nodes. Doing so, one insures the correct representation of the near null space of the
problem on coarser levels.
Note that, alike classical multigrid methods based on coarsening by aggregation
[11, 20, 48], Reitzinger and Schoberl (RS) approach has low computational cost per it-
eration and modest storage requirements. However, similarly to classical aggregation
techniques, piecewise constant (up to the edge’s orientation) prolongation is used, which
in turn results in level-dependent convergence behaviour with V-cycle setting (as already
observed in [52]).
Regarding aggregation techniques for elliptic boundary value problems, several ap-
proaches have been proposed recently to overcome this lack of optimality. One consists
in using, instead of simple V-cycle, a more sophisticated K-cycle, in which Krylov sub-
space acceleration is performed at each level [49]. It is also possible to improve the
scalability by increasing the number of smoothing steps on coarser levels [32]. Such
approaches can also be implemented with RS algebraic multigrid, keeping the original
advantage of modest resource requirements. As for the second implementation, numer-
ical experiments in the original RS paper [52] seem to indicate that this approach has
level-dependent convergence similar to that of V-cycle. Regarding the implementation
of RS approach with Krylov-based (K-) cycling strategy, it is however an open question
to what extent level-independent convergence properties can be obtained.
Here we investigate this point. We start by assessing the two-grid convergence of the
RS approach, since a (truly) multigrid method can not be optimal if the convergence rate
of the corresponding two-grid scheme deteriorates with the problem size. We evaluate
the convergence properties using Fourier analysis. This technique was adapted only
recently in [9] to curl-curl problems and, as far as we know, no such analysis is available
for the RS approach.
More precisely, a (local) two-grid Fourier analysis for a two-dimensional model prob-
lem based on geometrical (bilinear) prolongation operator is performed in [9] for hy-
brid [29] and AFW block [2] smoothers. Here we extend presented ideas to a piecewise
constant (RS-like) edge prolongation and show that the considered two-grid scheme with
hybrid smoother has level independent convergence properties in three dimensions. The
use of three-dimensional setting is motivated by its importance in the field of electro-
magnetical computations.
Fourier Analysis of aggregation-based two-grid method for edge element 115
Note that instead of using the local Fourier analysis framework, we perform an ex-
act Fourier analysis for problems with periodic boundary conditions. Both approaches
are quite similar, the latter allowing however to account for the grid size. This further
enables to supplement the Fourier analysis results with the assessment of convergence
properties of the two-grid scheme for the same problem with Dirichlet boundary condi-
tions. Their comparison indicates that Fourier analysis gives an accurate prediction of
the convergence rate.
Once the level-independent convergence is proved for a two-grid scheme, some nu-
merical experiments are performed using the corresponding multigrid ingredients with
V-, W- and K-cycles (the two latter approaches have similar operation count per itera-
tion). The results indicate that the convergence speed in the case of the first two cycling
strategies deteriorates with the number of levels, whereas the last approach has almost
the same iteration count as the two-grid scheme on the finest level.
The reminder of this paper is organized as follows. In Section 6.2 we recall some
useful properties of discrete curl-curl problems and present the main ingredients of the
RS approach. In Section 6.3 we give the Fourier representation of these ingredients for
the considered three-dimensional model problem. The results of the Fourier analysis
together with numerical experiments are presented and discussed in Section 6.4.
6.2 Preliminaries
6.2.1 Discretized problem
The use of edge finite elements requires the weak formulation of the problem (6.1) as
can be found, for instance, in [29]. More precisely, letting
H(curl; Ω) =v ∈ L2(Ω); curl(v) ∈ L2(Ω)
,
the “weak” problem consists in determining the vector E ∈ H∗(curl; Ω) ⊂ H(curl; Ω)
such that∫Ωα curl(E) · curl(v)dV +
∫ΩβE · vdV =
∫Ω
f · vdV ∀v ∈ H∗(curl; Ω) . (6.2)
This formulation can be recovered from the original problem, assuming that∫∂Ω
(curl(E)× v) · n dσ =∫∂Ω
curl(E) · (v × n) dσ = 0 (6.3)
116 Fourier Analysis of aggregation-based two-grid method for edge element
holds, through the multiplication of (6.1) by a test function v followed by the application
of Green’s identity∫Ω
curl(w) · vdV −∫
Ωw · curl(v)dV =
∫∂Ω
(w × v)ndσ .
The condition (6.3) is fulfilled, for instance, when homogeneous Dirichlet boundary
conditions
v × n = 0 ∀v ∈ H∗(curl; Ω)) = H0(curl; Ω)) (6.4)
are used.
In edge element discretization the degrees of freedom are associated with edges; that
is, to any edge denoted by k = (j1, j2) with nodes j1 and j2 being, respectively, the
starting and the end points, corresponds an unknown given by
xk =∫ j2
j1
E · ds .
The resulting system
Ax = b
is then such that
A = αKcc + βh2M , (6.5)
where Kcc and M are matrices that correspond, respectively, to the stiffness (curl-curl)
and the mass terms in (6.2).
One of the reasons why classical multigrid does not suit for such problems is the large
near null space of A induced by the null space N (Kcc) of Kcc. This latter is a discrete
representation of the null space of curl(·) operator, which contains all vectors of the form
grad(f). That is, N (Kcc) is formed by the vectors Gv, where G is a discrete gradient
matrix. As a straightforward consequence, we thus have KccG = O. When edge shape
functions are properly normalized, it can be proved (see [8] and the references therein)
that G coincides with the edge-node incidence matrix; that is, denoting by (j1, j2) an
edge with j1 as a starting node and j2 as the end node, we have
(G)kj =
1 if k = (∗, j) ,−1 if k = (j, ∗) ,
0 otherwise .
(6.6)
Since G associates a given node with several edges, it can be viewed as a transfer
operator from nodal to edge representation, its transpose GT performing the inverse
operation. Note, however, that the number of nodes and edges is generally not the same
and GTG, GGT 6= I.
Fourier Analysis of aggregation-based two-grid method for edge element 117
6.2.2 Reitzinger and Schoberl (RS) multigrid
It is a common practice to base the design of an algebraic multigrid method on the
definition of a problem-dependent prolongation matrix P . Once the prolongation is
available, the restriction is set to its transpose P T and the coarse grid matrix is given
by the Galerkin formula Ac = P TAP . The same procedure can then be applied to the
coarse grid system, and so on, until the coarsest grid which is chosen small enough.
The above ideas can be extended to algebraic multigrid for edge element discretizations
of (6.2), provided that they are combined with an appropriate smoothing scheme (for
instance, hybrid [29] or AFW [2] smoothers).
Now, in the RS approach, one first performs an agglomeration of n nodes into nc > 0
aggregates Γk, k = 1, ..., nc. The edge prolongation matrix is then defined by
(P (e))jk =
1 if j = (j1, j2) and k = (k1, k2) with j1 ∈ Γk1 , j2 ∈ Γk2−1 if j = (j1, j2) and k = (k2, k1) with j1 ∈ Γk1 , j2 ∈ Γk2
0 otherwise .
(6.7)
In other words, edges are grouped together in a unique “edge” aggregate if they connect
nodes belonging to same “node” aggregates.
Note that, setting the auxiliary “nodal” prolongation to
(P )(n)jk =
1 if j ∈ Γk, k = 1, ..., nc ,
0 otherwise ,(6.8)
one satisfies a seemingly important commutation property (see [52] for the proof)
GP (n) = P (e)Gc , (6.9)
with G and Gc being, respectively, the fine and coarse edge-node incidence matrices.
The importance of the property (6.9) mainly resides in the fact that the columns of G
span the near null space of A. The commutation property then ensures that the columns
of Gc belong to the near null space of Ac = P (e) T AP (e).
We also observe that the range of the prolongation matrix P (e) as defined by (6.7)
does not contain the entirety of the near null space of A. Therefore, some near null space
components of the error are not reduced appropriately by the coarse-grid correction. On
the other hand, a simple pointwise smoother R cannot reduce these components as well,
since (I − R−1A)v ≈ (1−O(h2))v for any v ∈ N (Kcc). More sophisticated smoothers
should therefore be used which treat appropriately the near null space modes that are
not in the range of the prolongation.
Here we consider one of such approaches known as the hybrid smoother. Its main
idea is to smooth separately the near null space components of the error, so that they
118 Fourier Analysis of aggregation-based two-grid method for edge element
can be correctly approximated on the coarser grid. The hybrid smoother involves two
additional matrices: an edges R(e) and a nodal R(n) smoother. If these matrices are
chosen to be the lower (or upper) triangular part of, respectively, A and A(n) = GTAG,
we recover the classical version of the hybrid smoother. In what follows we however
also consider diagonal (or Jacobi-like) smoothers. In the smoothing procedure given
below, the number of smoothing steps can be integer or half-integer. In the former case
an additional binary parameter η is used to determine if the extra half-step should be
performed in the beginning (η = ↑) or at the end (η = ↓) of the hybrid smoothing
scheme.
Hybrid smoother: xn+1 = HS(xn,b, ν, η)
(1) if ν is half-integer and η = ↑ : perform the steps (b), (d)-(f) below
(2) repeat bνc times:
(a) Edge pre-smoothing: xn ← xn + R(e)−1(b−Axn)
(b) Restrict to nodal variables: r = GT (b−Axn); e = 0
(c) Forward sweep: e← e + R(n)−1 (r−A(n)e
)(d) Backward sweep: e← e + R(n)−T (r−A(n)e
)(e) Transfer back to edge variables: xn ← xn +Ge
(f) Edge post-smoothing: xn+1 ← xn + R(e)−T (b−Axn)
(3) if ν is half-integer and η = ↓ : perform the steps (a)-(c), (e) above
Now, the two-grid version of the RS approach based on the hybrid smoother is presented
below. Note that the pre- and post-smoothing setting are treated differently when the
number of smoothing steps is half-integer. The reason for doing so is that the resulting
two-grid (and the induced multigrid) preconditioner is then symmetric when ν1 = ν2.
RS two-grid cycle: xn+1 = RSTG(b,xn, ν1, ν2)
(1) ν1 steps of pre-smoothing : xn ← HS(xn, ν1,b, ↑ )
(2) Compute residual: r = b−Axn(3) Restrict residual: rc = P (e) T r
(4) Coarse grid correction: ec = A−1c rc
(5) Prolongate coarse-grid correction: xn ← xn + P (e)ec(6) ν2 steps of post-smoothing : xn+1 ← HS(xn, ν2,b, ↓ )
When applying this algorithm, the error satisfies
A−1b− xn+1 = ETG(A−1b− xn
),
where the iteration matrix ETG is given by
E(ν1,ν2)TG = S
(ν2)↓
(I − P (e)
(P (e) T AP (e)
)−1P (e) T A
)S
(ν1)↑ . (6.10)
Fourier Analysis of aggregation-based two-grid method for edge element 119
The pre-smoothing iteration matrix satisfies
S(ν1)↑ =
S ν1 if ν1 is integer
S ν1 (I − R(e)−T A)(I −G R(n)−T GTA) if ν1 is half-integer ,(6.11)
where
S = (I − R(e)−T A)(I −G X(n)−1GTA)(I − R(e)−1
A) , (6.12)
with X(n) defined by
I − X(n)−1A(n) = (I − R(n)−T A(n))(I − R(n)−1
A(n)) . (6.13)
Similarly, the post-smoothing iteration matrix is given by
S(ν2)↓ =
S ν2 if ν2 is integer
(I −G R(n)−1GTA)(I − R(e)−1
A) S ν2 if ν2 is half-integer ,(6.14)
Our main objective is the analysis of the spectral radius ρ (ETG) of ETG , which
governs convergence of the two-grid method. We note, however, that the RS multigrid
method can be used as a preconditioner, with the preconditioner matrix BTG in the
two-grid case given by
ETG = I −BTGA .
Since the system matrix A resulting from the edge element discretization of (6.2) is
symmetric positive definite (SPD), and since BTG can be checked to be SPD for ν1 = ν2,
the linear system can be solved by the preconditioned conjugated gradient method [28].
In this latter case, the relevant convergence parameter is the condition number (see,
e.g., [24, Theorem 10.2.6]), given by
κ (BTGA) =λmax (BTGA)λmin (BTGA)
=1− λmin (ETG)1− λmax (ETG)
.
Unless the coarse grid matrix is weighted (as it is sometimes the case below), one can
check that A1/2ETGA−1/2 = I − A1/2BTGA
1/2 is semi-positive definite (see Theorem
3.19 in [65] for nonnegative definiteness of BTG−1 − A and Theorem 2.1 in [44] with
nc > 0 for presence of zero eigenvalues) and, hence, λmin(ETG) = 0. The condition
number in such case can therefore be deduced from
κ (BTGA) =1
1− ρ (ETG), (6.15)
and will not be reported explicitly.
120 Fourier Analysis of aggregation-based two-grid method for edge element
Note that, since AS(ν)↑ = S
(ν)↓
TA and AC = CTA, where C stands for the coarse
grid correction, we have
ρ(E
(ν1,ν2)TG
)= ρ
(E
(ν1,ν2)TG
T)
= ρ
(A−1E
(ν1,ν2)TG
TA
)= ρ
(A−1 S
(ν1)↑
TC T S
(ν2)↓
TA
)= ρ
(S
(ν1)↓ CS
(ν2)↑
)= ρ
(E
(ν2,ν1)TG
), (6.16)
and the number of pre- and post-smoothing iterations can be interchanged without any
impact on the asymptotic two-grid convergence. Moreover, if ν1 and ν2 are both integers
or half-integers, using A(n) = GTAG we have S(1/2)↑ S
(1/2)↓ = S, which further implies
S(ν1)↑ S
(ν2)↓ = Sν1+ν2 = S
(ν1+ν2)↑ = S
(ν1+ν2)↓ ,
and, hence,
ρ(E
(ν1,ν2)TG
)= ρ
(S
(ν2)↓ CS
(ν1)↑
)= ρ
(CS
(ν1)↑ S
(ν2)↓
)= ρ
(CS
(ν1+ν2)↑
)= ρ
(E
(ν1+ν2,0)TG
).
In this case the two-grid convergence factor depends only on the overall number ν =
ν1 + ν2 of smoothing steps.
Now, in what follows we consider the hybrid smoother with ν = 1/2, 1 and 2 smooth-
ing iterations. In the two latter cases both ν1 and ν2 are either integer or half-integer;
hence, the asymptotic convergence factor then depends only on ν. The case ν = 1/2
corresponds to either (ν1, ν2) = (1/2, 0) or (0, 1/2). However, it follows from (6.16)
that both have the same asymptotic convergence, this latter depending again on ν. We
therefore report the results with respect to ν instead of (ν1, ν2), at least in the two-grid
setting.
6.3 Fourier analysis
6.3.1 Model problem
Consider now Ω = (0, 1)3 with periodic boundary conditions. The vectors in
H∗(curl; Ω) = HP (curl; Ω) are therefore also assumed periodic; that is, for any v ∈HP (curl; Ω) we have v(0, y, z) = v(1, y, z) , v(x, 0, z) = v(x, 1, z) and v(x, y, 0) =
v(x, y, 1) . Note that the constraint (6.3) is then satisfied since the contributions of
Fourier Analysis of aggregation-based two-grid method for edge element 121
opposite faces of ∂Ω are opposite. The weak formulation (6.2) can therefore be consid-
ered and the resulting problem is further discretized by trilinear (brick) edge elements1
(see, e.g., [66, p.54]) on the cubic grid (N + 1)× (N + 1)× (N + 1) of grid size h = N−1.
Since it is sometimes convenient to refer to an edge via its position on the grid, we
also associate a triple k = (kx, ky, kz) to any node unknown such that hk gives node’s
coordinate position, and to any edge unknown such that hk correspond to coordinate
position of the corresponding edge’s middle point. Note that ka, a = x, y, z, is a half-
integer if the corresponding edge is oriented in the a direction and integer otherwise.
Now, following the notation in [9], we set
I(∆x,∆y,∆z) = (kx + ∆x, ky + ∆y, kz + ∆z|0 ≤ kx, ky, kz < N) ,
and let E [x] = I(1/2, 0, 0), E [y] = I(0, 1/2, 0), E [z] = I(0, 0, 1/2) and N = I(0, 0, 0) be
the index set of, respectively, edge unknowns in x, y and z directions and node unknowns.
We also note that, for any edge k, the set of its neighbours; that is, the set of edges that
have a common element with k is given by 〈k + t〉 = (〈kx + tx〉, 〈ky + ty〉, 〈kz + tz〉)T ,
where t ∈ T , with
T = t = (tx, ty, tz) | tx, ty, tz ∈ 1,12, 0, −1
2, −1 and tx + ty + tz ∈ Z
and
〈k〉 =
k if k < N ,
k −N otherwise .
Assuming that the matrix A arises from the discretization of (6.2) with coefficients
α and β being constant, the entry (A)kk′ for a given edge orientation depends on the
relative edge’s position k− k′, and, hence, satisfies
(Av)k =
∑
t∈T s[x]t (v)〈k+t〉 if k ∈ E [x] ,∑
t∈T s[y]t (v)〈k+t〉 if k ∈ E [y] ,∑
t∈T s[z]t (v)〈k+t〉 if k ∈ E [z] .
(6.17)
Similarly to the two-dimensional analysis in [9], we associate a stencil to edges in any ofthe three directions. For instance, for edges in x direction the stencil can be represented
1the elements that belong to the boundary edges being periodically extended to the opposite bound-ary.
122 Fourier Analysis of aggregation-based two-grid method for edge element
f ff
f f
fvf
ff
ff fff
vf
ff
Figure 6.1: Edge neighbourhood.
as a triple of two-dimensional stencils
ST [x](L) =
s[x]
− 12 ,−
12 ,1
s[x]
− 12 ,
12 ,1
s[x]
− 12 ,−1, 12
s[x]
− 12 ,0,
12
s[x]
− 12 ,1,
12
s[x]
− 12 ,−
12 ,0
• s[x]
− 12 ,
12 ,0
s[x]
− 12 ,−1,− 1
2s[x]
− 12 ,0,−
12
s[x]
− 12 ,1,−
12
s[x]
− 12 ,−
12 ,−1,
s[x]
− 12 ,
12 ,−1
s[x]0,−1,1 s
[x]0,0,1 s
[x]0,1,1
s[x]0,−1,0 s
[x]0,0,0 s
[x]0,1,0
s[x]0,−1,−1 s
[x]0,0,−1 s
[x]0,1,−1
s[x]12 ,−
12 ,1
s[x]12 ,
12 ,1
s[x]12 ,−1, 12
s[x]12 ,0,
12
s[x]12 ,1,
12
s[x]12 ,−
12 ,0
• s[x]12 ,
12 ,0
s[x]12 ,−1,− 1
2s[x]12 ,0,−
12
s[x]12 ,1,−
12
s[x]12 ,−
12 ,−1,
s[x]12 ,
12 ,−1
,
the edges in this stencil being also represented on Figure 6.1. More particularly, the
“bold” segment corresponds to the considered edge and to the entry s[x]0,0,0, the other 8
edges in x direction forming the rest of the central 2D stencil; the remaining 24 edges are
oriented in y and z direction and belong to two planes, those with smaller x coordinate
corresponding to the first 2D stencil, the others being associated to the third one. For
these two stencils, the black and white bullets schematize the nodes (as on Figure 6.1)
and the value between two bullets in the stencil corresponds to the edge between them
on the figure.
Note that the same stencil representation can be used in y and z directions. In
these latter cases, to avoid any confusion on the choice of directions perpendicular to
the considered edge, we assume that stencil entries s[a]tx,ty ,tz , a = y, z, of every column
have the same tx, and the entries of every line have the same tz. Note that the stencil
in the x direction given above also satisfies this assumption.
Fourier Analysis of aggregation-based two-grid method for edge element 123
Now, assuming edge elements oriented in the positive axis direction, the matricesKcc and M for the considered problem satisfy (6.17) with
ST [x](Kcc) =16
1 −1 −1 −4 −1
4 • −4 1 4 1
1 −1
−2 −2 −2
−2 16 −2
−2 −2 −2
−1 1 1 4 1
−4 • 4 −1 −4 −1
−1 1
,
(6.18)
and
ST [x](M) =136
[•]
1 4 1
4 16 4
1 4 1
[•], (6.19)
respectively. The stencils are the same in the y and z directions.
6.3.2 Fourier analysis setting
For edge unknowns, we define the Fourier modes separately in each direction:
(u[a](θ)
)k
=
eikθ = 1√
N3ei(kxθx+kyθy+kzθz) if k ∈ E [a]
0 otherwise ,(6.20)
whereas for the node unknowns the usual definition is adopted
(u(θ))k = eikθ =1√N3
ei(kxθx+kyθy+kzθz) , k ∈ N . (6.21)
The following abbreviations are used in the rest of the chapter:
ca = cos(θa/2) and sa = sin(θa/2) , a = x, y, z . (6.22)
The proposition below shows that the subspace spanned by a triple of edge modes(u[x](θ) u[y](θ) u[z](θ)
)is invariant with respect to the system matrix A if θ ∈ Θ, where
Θ =
(2π`xN
,2π`yN
,2π`zN
)T| `x, `y, `z ∈ N and 0 ≤ `x, `y, `z < N
. (6.23)
That is, in Fourier basis we have A = diag(A(θ)), with A(θ) given by (6.25).
Proposition 6.1. Let A be defined by (6.5), where Kcc and M are edge matrices on
a (N + 1) × (N + 1) × (N + 1) cubic grid satisfying (6.17) with stencils in x, y and
124 Fourier Analysis of aggregation-based two-grid method for edge element
z directions given by (6.18) and (6.19), respectively. Let u[a](θ), a = x, y, z and Θ be
defined by (6.20) and (6.23), respectively.
Then, for any θ ∈ Θ there holds
A(u[x](θ) u[y](θ) u[z](θ)
)=(u[x](θ) u[y](θ) u[z](θ)
)A(θ) (6.24)
where
A(θ) = αKcc(θ) + βh2M(θ) (6.25)
with
Kcc(θ) =43
3s2y + 3s2
z − 4s2ys
2z −sxsy(3− 2s2
z) −sx(3− 2s2y)sz
−sxsy(3− 2s2z) 3s2
x + 3s2z − 4s2
xs2z −(3− 2s2
x)sysz−sx(3− 2s2
y)sz −(3− 2s2x)sysz 3s2
x + 3s2y − 4s2
xs2y
(6.26)
M(θ) =19
(3− 2s2
y)(3− 2s2z)
(3− 2s2x)(3− 2s2
z)
(3− 2s2x)(3− 2s2
y)
(6.27)
and with sa, a = x, y, z , given by (6.22).
Proof. Note that, using (6.17), (6.18) and (6.19), we have
Au[z](θ) = α13
(8− eiθx − e−iθx − eiθy − e−iθy
− ei(θx+θy) − ei(−θx+θy) − ei(θx−θy) − e−i(θx+θy))u[x](θ)
+ α16
(e−iθz/2(4e−iθy/2 − 4eiθy/2 − ei(θy/2+θx) + ei(−θy/2+θx) − ei(θy/2−θx) + ei(−θy/2−θx))
+ eiθz/2(4eiθy/2 − 4e−iθy/2 + ei(θy/2+θx) − ei(−θy/2+θx) + ei(θy/2−θx) − ei(−θy/2−θx)))u[y](θ)
+ α16
(e−iθz/2(4e−iθx/2 − 4eiθx/2 − ei(θx/2+θy) + ei(−θx/2+θy) − ei(θx/2−θy) + ei(−θx/2−θy))
+ eiθz/2(4eiθx/2 − 4e−iθx/2 + ei(θx/2+θy) − ei(−θx/2+θy) + ei(θx/2−θy) − ei(−θx/2−θy)))u[z](θ)
+ β136h2(
16 + 4eiθx + 4e−iθx + 4eiθy + 4e−iθy
+ +ei(θx+θy) + ei(−θx+θy) + ei(θx−θy) + e−i(θx+θy))u[x](θ)
Fourier Analysis of aggregation-based two-grid method for edge element 125
and, after some tedious trigonometry, the last column of (6.26) and (6.27) follows. The
other lines are determined similarly.
Regarding the edge-node incidence matrix, the general expression (6.6) can be further
rewritten for the considered grid as
(G)kekn =
1 if kn = 〈ke+(1
2 , 0, 0)〉 , 〈ke+(0, 12 , 0)〉 or 〈ke+(0, 0, 1
2)〉−1 if kn = 〈ke−(1
2 , 0, 0)〉 , 〈ke−(0, 12 , 0)〉 or 〈ke−(0, 0, 1
2)〉0 otherwise .
(6.28)
The following theorem gives the Fourier representation of G.
Proposition 6.2. Let G be defined by (6.28) on a (N+1)×(N+1)×(N+1) cubic grid.
Let u[a](θ), a = x, y, z, u(θ) and Θ be defined by (6.20), (6.21) and (6.23), respectively.
Then, for any θ ∈ Θ there holds
GT(u[x](θ) u[y](θ) u[z](θ)
)= 2iu(θ)(sx sy sz) (6.29)
and with sa, a = x, y, z , given by (6.22).
Proof. Note that for any k ∈ N there holds
(GTu[x](θ))k =1√N3
(ei((kx+1/2)θx+kyθy+kzθz) − ei((kx−1/2)θx+kyθy+kzθz)
)=(eiθx/2 − e−iθx/2
)(u(θ))k
= 2i sin (θx/2) (u(θ))k ,
which gives the first entry of (6.29). The proof for the other entries is similar.
Now, we assume that N is even and consider two types of aggregation patterns:
(xy) we aggregate nodes into squares in xy-plane, leading to
Γxyk = (2kx, 2ky, kz), (2kx+1, 2ky, kz), (2kx, 2ky+1, kz), (2kx+1, 2ky+1, kz)
with kx, ky, kz being integer and such that 0 ≤ 2kx, 2ky, kz < N .
(xyz) we aggregate nodes into cubes by grouping the nodes
Γxyzk = (2kx, 2ky, 2kz), (2kx+1, 2ky, 2kz), (2kx, 2ky+1, 2kz), (2kx+1, 2ky+1, 2kz),
(2kx, 2ky, 2kz+1), (2kx+1, 2ky, 2kz+1), (2kx, 2ky+1, 2kz+1), (2kx+1, 2ky+1, 2kz+1) ,
with kx, ky, kz being integer and such that 0 ≤ 2kx, 2ky, 2kz < N .
126 Fourier Analysis of aggregation-based two-grid method for edge element
We extend our coordinate notation to the coarse grid edge unknowns, letting
I`(∆x,∆y,∆z) =
(kx+∆x, ky+∆y, kz+∆z|0 ≤ 2kx, 2ky, kz < N) if ` = xy ,
(kx+∆x, ky+∆y, kz+∆z|0 ≤ 2kx, 2ky, 2kz < N) if ` = xyz ,
and setting their index set to E [x]` = I`(1/2, 0, 0), E [y]
` = I [z]` (0, 1/2, 0) and E [z]
` =
I`(0, 0, 1/2), ` = xy, xyz, for edges oriented in x, y and z direction, respectively.The edge prolongation defined by (6.7) is then given for any k = (kx, ky, kz) by
(P (e)xy
Tw)k =
(w)k1 + (w)k2 ,
k1 =(2kx+1/2, 2ky+1, kz),k2 =(2kx+1/2, 2ky, kz) if k ∈ E [x]xy ,
(w)k1 + (w)k2 ,
k1 =(2kx+1, 2ky+1/2, kz), k2 =(2kx, 2ky+1/2, kz) if k ∈ E [y]xy ,
(w)k1 + (w)k2 + (w)k3 + (w)k4 ,
k1 =(2kx+1, 2ky+1, kz) ,k2 =(2kx+1, 2ky, kz) ,
k3 =(2kx, 2ky+1, kz) ,k4 =(2kx, 2ky, kz) if k ∈ E [z]xy .
(6.30)
in (xy) case and by
(P (e)xyz
Tw)k =
(w)k1 + (w)k2 + (w)k3 + (w)k4 ,
k1 =(2kx+1/2, 2ky+1, 2kz+1), k2 =(2kx+1/2, 2ky, 2kz),
k3 =(2kx+1/2, 2ky, 2kz+1), k4 =(2kx+1/2, 2ky+1, 2kz) if k ∈ E [x]xyz ,
(w)k1 + (w)k2 + (w)k3 + (w)k4 ,
k1 =(2kx+1, 2ky+1/2, 2kz+1), k2 =(2kx, 2ky+1/2, 2kz),
k3 =(2kx, 2ky+1/2, 2kz+1), k4 =(2kx+1, 2ky+1/2, 2kz) if k ∈ E [y]xyz ,
(w)k1 + (w)k2 + (w)k3 + (w)k4 ,
k1 =(2kx+1, 2ky+1, 2kz+1/2), k2 =(2kx, 2ky, 2kz+1/2),
k3 =(2kx, 2ky+1, 2kz+1/2), k4 =(2kx+1, 2ky, 2kz+1/2) if k ∈ E [z]xyz
(6.31)
in (xyz) case. Further, we define the coarse grid Fourier modes for ` = xy, xyz, and
a = x, y, z, as
(u[a]` (θ)
)k
=
eikθ = 1√
N3ei(kxθx+kyθy+kzθz) if k ∈ E [a]
`
0 otherwise .(6.32)
As shown in the following proposition, the frequency aliasing is then such that all Fouriermodes (6.20) corresponding to the frequencies in Θ`(θ), ` = xy, xyz, lead to a uniquefrequency on the coarse grid, with
Θxy(θ) =((θx, θy, θz)T , (θx + π, θy, θz)T , (θx, θy + π, θz)T , (θx + π, θy + π, θz)T
), (6.33)
and
Θxyz(θ) =((θx, θy, θz)T , (θx+π, θy, θz)T , (θx, θy+π, θz)T , (θx+π, θy+π, θz)T
Fourier Analysis of aggregation-based two-grid method for edge element 127
(θx, θy, θz+π)T , (θx+π, θy, θz+π)T , (θx, θy+π, θz+π)T , (θx+π, θy+π, θz+π)T). (6.34)
Proposition 6.3. Let P (e)xy , Θxy and P
(e)xyz, Θxyz be defined by (6.30), (6.33) and by
(6.31), (6.34), respectively. Let u[a](θ) and u[a]` (θ), a = x, y, z, ` = xy, xyz, be defined
by (6.20) and (6.32) and set sa and ca as in (6.22).Then, for any (θ1, θ2, θ3, θ4) = Θxy(2π`x
N ,2π`yN , 2π`z
N ), `x, `y, `z ∈ N, and for anya = x, y, z , there holds
P (e)xy
T(u[a](θ1) u[a](θ2) u[a](θ3) u[a](θ4)
)= u[a]
xy(θc) Pxy(θc)H,
where θc =(
22π`xN , 22π`y
N , 2π`zN
)Tand
Pxy(θc)H =
−2ei(θx+θy)/2 (−cy icy isy sy ) if a = x ,
−2ei(θx+θy)/2 (−cx isx icx sx ) if a = y ,
−4ei(θx+θy)/2 (−cycx icysx isycx sysx ) if a = z .
Similarly, for any (θ1, θ2, θ3, θ4, θ5, θ6, θ7, θ8) = Θxyz(2π`xN ,
2π`yN , 2π`z
N ), `x, `y, `z ∈ N,and for any a = x, y, z , there holds
P (e)xyz
T(u[a](θ1) u[a](θ2) u[a](θ3) u[a](θ4) u[a](θ5) u[a](θ6) u[a](θ7) u[a](θ8)
)= u[a]
xyz(θc) P [a]xyz(θc)
H,
where θc =(
22π`xN , 22π`y
N , 22π`zN
)Tand
P [a]xyz(θc)
H=
−4ei(θx+θy+θz)/2(−cycz icycz isycz sycz icysz cysz sysz −isysz ) if a = x ,
−4ei(θx+θy+θz)/2(−cxcz isxcz icxcz sxcz icxsz sxsz cxsz −isxsz ) if a = y ,
−4ei(θx+θy+θz)/2(−cxcy isxcy icxsy sxsy icxcy sxcy cxsy −isxsy ) if a = z .
Proof. We indicate the proof for P (e)xy when a = x, the proof is similar in the other
cases. For any k ∈ E [x]xy , setting θ1 = (θx, θy, θz), we have(
P (e)xy
Tu[x](θ1)
)k
= ei((2kx+1/2)θx+(2ky+1)θy+kzθz) + ei((2kx+1/2)θx+2kyθy+kzθz)
= eikθceiθx/2(1 + eiθy)(P (e)xy
Tu[x](θ2)
)k
= ei((2kx+1/2)(θx+π)+(2ky+1)θy+kzθz) + ei((2kx+1/2)(θx+π)+2kyθy+kzθz)
= eikθcei2kxπei(θx+π)/2(1 + eiθy)(P (e)xy
Tu[x](θ3)
)k
= ei((2kx+1/2)θx+(2ky+1)(θy+π)+kzθz) + ei((2kx+1/2)θx+2ky(θy+π)+kzθz)
= eikθceiθx/2(1− eiθy)(P (e)xy
Tu[x](θ4)
)k
= ei((2kx+1/2)(θx+π)+(2ky+1)(θy+π)+kzθz) + ei((2kx+1/2)(θx+π)+2kyθy+kzθz+)
= eikθcei2kxπei(θx+π)/2(1− eiθy)
128 Fourier Analysis of aggregation-based two-grid method for edge element
and, since 1+eiθy = cyeiθy/2, 1−eiθy = −isyeiθy/2, eiπ/2 = i and ei2kxπ = −1, the desired
result follows.
Now, note that the Propositions 6.1 and 6.3 imply that both the system matrix A
and the coarse grid correction matrix I − P (e)`
(P
(e)`
TAP
(e)`
)−1
P(e)`
TA, ` = xy, xyz ,
with P` given either by (6.30) or by (6.31), possess invariant subspacesu[a](θ) | a = x, y, z and θ ∈ Θ`(
2π`xN
,2π`yN
,2π`zN
)
(6.35)
with `x, `y, `z being integer. Since the union of such subspaces for 0 ≤ `x, `y, `z < N
forms an orthogonal basis in R3N3, the coarse grid correction matrix has a block diagonal
structure in such basis with m×m blocks (m = 12 in the (xy) case and m = 24 if (xyz)
is considered). If, in addition, the subspace (6.35) is invariant under the smoothing
iteration matrix (6.14), the same conclusion on the block structure holds for the two-
grid iteration matrix (6.10); that is, in the Fourier basis ETG = diag(Ξ`x,`y ,`z
)with
Ξ`x,`y ,`z being a m×m matrix. The invariance requirement on the smoothing iteration
matrix is in turn fulfilled if there exist matrices R(e) and R(n) such that
R(e)(u[x](θ) u[y](θ) u[z](θ)
)=(u[x](θ) u[y](θ) u[z](θ)
)R(e)(θ) , (6.36)
R(n)u(θ) = u(θ)R(n)(θ) . (6.37)
It is then possible to assess the two-grid convergence factor via the spectral radii of
Ξ`x,`y ,`z , namely
ρ (ETG) = max`x,`y ,`z
ρ(Ξ`x,`y ,`z
).
If both R(e) and R(n) are of Jacobi type; that is, if
R(e) =1ω(e)
diag(A) , (6.38)
R(n) =1ω(n)
diag(A(n)) , (6.39)
then (6.36) and (6.37) hold with
R(e)(θ) =1ω(e)
(α
166
+ βh2 1636
)Im ,
R(n)(θ) =1ω(n)
diag(A(n)) =1ω(n)
diag(GTMG) =1ω(n)
βh2 83,
the second equality of the second line coming from KccG = O.
If R(e) and R(n) are of Gauss-Seidel type; that is, if R(e) and R(n) are (up to some
reordering of unknowns) upper (or lower) triangular part of A and GTAG, respectively,
then the relations (6.36) and (6.37) are not satisfied. However, following the usual
Fourier Analysis of aggregation-based two-grid method for edge element 129
practice [61, 68, 8], we can approximate these matrices by R(e) and R(n) which satisfy
(6.36) and (6.37), respectively.
In particular, since A(n) = GTMG has the following three-dimensional stencil−1 −2 −1
−2 0 −2
−1 −2 −1
−2 0 −2
0 32 0
−2 0 −2
−1 −2 −1
−2 0 −2
−1 −2 −1
,
which also correspond to a trilinear discretization of Poisson equation, the stencil of R(n)
can be chosen as
ST (R(n)) =112
[ · ]
0 0 0
0 32 0
−2 0 −2
−1 −2 −1
−2 0 −2
−1 −2 −1
,
if the nodal unknowns are updated in lexicographical order. Then,
R(n)(θ) =112
(32− 4e−iθz(3− 3s2
x − 3s2y + 4s2
xs2y
)− 4e−iθy(1− 2s2
x)) ,
with θ = (θx, θy, θz) ∈ Θ.
For the edge Gauss-Seidel smoother, different strategies can be considered, depending
on the order in which the edge unknowns are updated.
direction-based strategy: edges in x direction are updated before those in y direction,
which in turn are updated before those in z direction; the ordering inside each direction
is lexicographical.
point-based strategy: edges that are associated to a particular node are updated
one after another (if not associated to an already updated node; that is, if not already
updated) ; the nodes are considered in lexicographical order.The first strategy can by approximated by the stencil
ST [x](R(e)) =16α
[•]
0 0 0
−2 16 0
−2 −2 −2
[•]
+136βh2
[•]
0 0 0
4 16 0
1 4 1
[•],
130 Fourier Analysis of aggregation-based two-grid method for edge element
in x direction, by
ST [y](R(e)) =16α
1 −1 0 0 0
4 • −4 0 0 0
1 −1
0 0 0
−2 16 0
−2 −2 −2
−1 1 0 0 0
−4 • 4 0 0 0
−1 1
+136βh2
[•]
0 0 0
4 16 0
1 4 1
[•],
in y direction and by
ST [z](R(e)) =16α
1 −1 −1 −4 −1
4 • −4 1 4 1
1 −1
0 0 0
−2 16 0
−2 −2 −2
−1 1 1 4 1
−4 • 4 −1 −4 −1
−1 1
+136βh2
[•]
0 0 0
4 16 0
1 4 1
[•],
in z direction. The corresponding Fourier block can be evaluated as in the proof of
Proposition 6.1, and is given by
(R(e)(θ))11 = α13
(8− e−iθy − e−iθz − 2(1− 2s2y)e−iθz) + βh2 1
9(4 + e−iθy + e−iθz +
12
(1− 2s2y)e−iθz)
(R(e)(θ))12 = 0
(R(e)(θ))13 = 0
(R(e)(θ))21 = −α43
sxsy(3− 2s2z)
(R(e)(θ))22 = α13
(8− e−iθx − e−iθz − 2(1− 2s2x)e−iθz) +
19βh2(4 + e−iθx + e−iθz +
12
(1− 2s2x)e−iθz)
(R(e)(θ))23 = 0
(R(e)(θ))31 = −α43
sx(3− 2s2y)sz
(R(e)(θ))32 = −α43
(3− 2s2x)sysz
(R(e)(θ))33 = α13
(8− e−iθx − e−iθy − 2(1− 2s2x)e−iθy) + βh2 1
9(4 + e−iθx + e−iθy +
12
(1− 2s2x)e−iθy) .
Fourier Analysis of aggregation-based two-grid method for edge element 131
The second strategy corresponds to the stencil
ST [x](R(e)) =16α
0 0 −1 0 0
4 • 0 1 4 1
1 −1
0 0 0
−2 16 0
−2 −2 −2
0 0 1 0 0
−4 • 0 −1 −4 −1
−1 1
+136βh2
[•]
0 0 0
4 16 0
1 4 1
[•],
in x direction, to the stencil
ST [y](R(e)) =16α
0 0 −1 0 0
4 • −4 1 4 1
1 −1
0 0 0
−2 16 0
−2 −2 −2
0 0 0 0 0
0 • 0 −1 −4 −1
−1 1
+136βh2
[•]
0 0 0
4 16 0
1 4 1
[•],
in y direction and to the stencil
ST [z](R(e)) =16α
0 0 −1 −4 0
4 • −4 1 4 1
1 −1
0 0 0
−2 16 0
−2 −2 −2
[•]
+136βh2
[•]
0 0 0
4 16 0
1 4 1
[•],
in z direction. The corresponding Fourier block is given by
(R(e)(θ))11 = α13
(8− e−iθy − e−iθz − 2(1− 2s2y)e−iθz) + βh2 1
9(4 + e−iθy + e−iθz +
12
(1− 2s2y)e−iθz)
(R(e)(θ))12 = αi
3sx(−4e−iθy/2 + 2isye−iθz)
132 Fourier Analysis of aggregation-based two-grid method for edge element
N ν = 1/2 ν = 1 ν = 250 xy. 0.852 0.794 0.755
xyz. 0.852 0.794 0.755100 xy. 0.853 0.796 0.759
xyz. 0.853 0.796 0.759150 xy. 0.853 0.796 0.760
xyz. 0.853 0.796 0.760
Table 6.1: Convergence factor of a two-grid method with Jacobi hybrid smoother,estimated via Fourier analysis.
point-based direction-basedN ν = 1/2 ν = 1 ν = 2 ν = 1/2 ν = 1 ν = 250 xy. 0.745 0.770 0.723 0.716 0.739 0.716
xyz. 0.770 0.777 0.726 0.716 0.740 0.716100 xy. 0.751 0.782 0.748 0.722 0.752 0.741
xyz. 0.776 0.789 0.751 0.722 0.753 0.741150 xy. 0.752 0.785 0.753 0.723 0.755 0.746
xyz. 0.777 0.792 0.756 0.723 0.755 0.746200 xy. 0.752 0.786 0.755 0.724 0.756 0.748
xyz. 0.777 0.793 0.758 0.724 0.756 0.748
Table 6.2: Convergence factor of a two-grid method with various Gauss-Seidel variantsof hybrid smoother, estimated via Fourier analysis.
(R(e)(θ))13 = αi
3sx(−(4 + eiθy)e−iθz/2 + 2isze−iθy)
(R(e)(θ))21 = α23
sx(−sye−iθz − ie−iθy/2)
(R(e)(θ))22 = α13
(8− e−iθx − e−iθz − 2(1− 2s2x)e−iθz) + βh2 1
9(4 + e−iθx + e−iθz +
12
(1− 2s2x)e−iθz)
(R(e)(θ))23 = α16
(−4isy(3− 2s2
x)e−iθz/2 − e−iθy/2e−iθxeiθz/2)
(R(e)(θ))31 = αi
3sxe−θz/2(−4− e−iθy)
(R(e)(θ))32 = α16e−iθz/2(eiθxe−iθy/2 − 2isy(4− e−iθx))
(R(e)(θ))33 = α13
(8− e−iθx − e−iθy − 2(1− 2s2x)e−iθy) + βh2 1
9(4 + e−iθx + e−iθy +
12
(1− 2s2x)e−iθy) .
6.4 Numerical results
6.4.1 Two-grid method
For the numerical investigations that follow, we set α = 1 and β = 0.01. When Jacobi
smoothers (6.38) and (6.39) are considered, the weights are chosen to be ω(e) = 1/3
and ω(n) = 2/3. This choice corresponds to the biggest values of weights such that the
iteration matrices I − R(e)−1A and I − R(n)−1
A are still positive definite for any N .
Fourier Analysis of aggregation-based two-grid method for edge element 133
ν = 1/2 ν = 1 ν = 2N FA D FA D FA D20 xy. 0.845 0.825 0.779 0.743 0.725 0.671
xyz. 0.845 0.844 0.779 0.775 0.725 0.72130 xy. 0.850 0.835 0.788 0.763 0.744 0.707
xyz. 0.850 0.849 0.788 0.786 0.744 0.74140 xy. 0.851 0.840 0.792 0.772 0.752 0.721
xyz. 0.851 0.851 0.792 0.791 0.752 0.749
Table 6.3: Comparison of Fourier analysis and actual two-grid convergence factors inthe case of Jacobi hybrid smoother.
ν = 1/2 ν = 1 ν = 2N FA P D FA P D FA P D20 xy. 0.711 0.640 0.600 0.689 0.683 0.646 0.573 0.587 0.583
xyz. 0.733 0.680 0.681 0.702 0.698 0.717 0.583 0.600 0.65930 xy. 0.733 0.675 0.639 0.741 0.746 0.693 0.667 0.678 0.636
xyz. 0.757 0.714 0.715 0.749 0.746 0.746 0.672 0.683 0.70640 xy. 0.741 0.687 0.662 0.760 0.750 0.715 0.705 0.711 0.672
xyz. 0.766 0.727 0.710 0.768 0.763 0.758 0.708 0.715 0.724
Table 6.4: Comparison of Fourier analysis and actual two-grid convergence factorsfor point-based Gauss-Seidel hybrid smoother.
We first consider Fourier analysis for large problem sizes. The corresponding results
are given in Table 6.1 for the Jacobi version of the hybrid smoother and in Table 6.2
for the different variants of its Gauss-Seidel version. The asymptotical values of the
convergence are (approximately) reached for N = 100 in the former case and for N = 150
in the latter. In both cases, the (almost) asymptotical values are bounded away from 1,
showing that RS approach has h-independent convergence properties in two-grid setting.
Note that periodic boundary conditions are rarely used in practice, their main pur-
pose here is to make the exact Fourier analysis possible. It is therefore instructive
to compare previous results with convergence factors of similar problems with realis-
tic boundary conditions. Here, the comparison is made with the problem (6.2) having
Dirichlet boundary conditions and discretized on the cubic grid (N+2)×(N+2)×(N+2)
of mesh size h = (N + 1)−1. Observe that, assuming the same number of unknowns,
this value of h differs slightly from the one defined in periodic case.
Now, the convergence factors for problems with periodic (FA) and Dirichlet (D)
boundary conditions are given in Table 6.3 for the Jacobi hybrid smoother. Tables 6.4
and 6.5 present the same information for smoothers of Gauss-Seidel type. In both cases,
we have evaluated real convergence factors using ARPACK [36] routines. Note that,
when a Gauss-Seidel smoother is considered, the corresponding matrices R(e) and R(n)
are approximated by R(e) and R(n) in order for Fourier analysis to be applicable. That
134 Fourier Analysis of aggregation-based two-grid method for edge element
ν = 1/2 ν = 1 ν = 2N FA P D FA P D FA P D20 xy. 0.673 0.662 0.602 0.479 0.642 0.599 0.560 0.538 0.537
xyz. 0.675 0.665 0.668 0.659 0.655 0.670 0.568 0.554 0.60830 xy. 0.701 0.698 0.652 0.709 0.702 0.652 0.658 0.651 0.603
xyz. 0.701 0.698 0.697 0.710 0.702 0.697 0.660 0.657 0.67940 xy. 0.711 0.709 0.671 0.730 0.726 0.685 0.697 0.694 0.643
xyz. 0.711 0.709 0.709 0.730 0.726 0.729 0.698 0.697 0.70750 xy. 0.716 0.715 0.682 0.739 0.739 0.702 0.716 0.714 0.670
xyz. 0.716 0.715 0.713 0.739 0.741 0.739 0.716 0.716 0.720
Table 6.5: Comparison of Fourier analysis and actual two-grid convergence factorsfor direction-based Gauss-Seidel hybrid smoother.
γ point-based GS direction-based GS Jacobiν = 1/2 ν = 1 ν = 2 ν = 1/2 ν = 1 ν = 2 ν = 1/2 ν = 1 ν = 2
1 0.777 0.792 0.756 0.723 0.755 0.746 0.853 0.796 0.7590.8 0.737 0.755 0.701 0.639 0.697 0.684 0.837 0.760 0.7040.6 0.686 0.706 0.614 0.472 0.605 0.581 0.819 0.712 0.6180.5 0.659 0.675 0.551 0.447 0.536 0.499 0.809 0.682 0.5550.4 0.644 0.642 0.376 0.594 0.444 0.376 0.798 0.667 0.4740.3 0.854 0.809 0.673 0.854 0.809 0.673 1.166 0.667 0.667
Table 6.6: Dependence of convergence rate on γ for (xyz) prolongation and varioussmoothers.
is why in this latter case the convergence assessed via Fourier analysis (FA) does not
coincide with the actual two-grid convergence for problem with periodic (P) boundary
conditions.
Regarding the values in these three tables, we observe that the Fourier analysis
seems to give an accurate, although sometimes pessimistic, estimate of the real two-
grid convergence. More generally, in view of all results presented so far it appears that
Gauss-Seidel implementations are superior to the Jacobi ones; the difference between
point-based and direction-based variants of Gauss-Seidel is small, the latter performing
globally better. We also observe that the use of more smoothing steps does not neces-
sarily pay off, and in some cases it can even slightly deteriorates the convergence; the
use of ν = 1/2 (or ν1, ν2 = 1/2 in case of symmetric multigrid method) seems to be
a good choice. Regarding the prolongations considered, the performance of (xy) and
(xyz) variants is similar, the latter being more attractive because of the faster decrease
in the size of coarser grid(s).
It is observed in [11] in the context of aggregation-based multigrid for Poisson-like
problems that a simple way to improve convergence is to use a weighted coarse grid
Fourier Analysis of aggregation-based two-grid method for edge element 135
γ point-based GS direction-based GS Jacobiν1, ν2 = 1/2 ν1, ν2 = 1 ν1, ν2 = 1/2 ν1, ν2 = 1 ν1, ν2 = 1/2 ν1, ν2 = 1
1 4.80 4.10 4.08 3.94 4.90 4.150.8 4.26 3.36 3.45 3.19 4.43 3.430.6 3.82 2.32 2.82 2.43 4.05 2.730.5 3.70 2.32 2.56 2.08 3.94 2.390.4 3.93 2.38 2.53 2.02 3.92 2.370.3 4.58 2.67 2.68 2.04 4.23 2.66
Table 6.7: Dependence of condition number on γ for (xyz) prolongation and varioussmoothers.
correction instead of the usual one. The same observations hold in the present edge-
based two-grid setting, replacing (6.10) by
ETG = S(ν2)↓
(I − γ−1P (e)
(P (e) T AP (e)
)−1P (e) T A
)S
(ν1)↑ ,
as can be seen in (xyz) case from Table 6.6. Since the relation (6.15) is not necessary
satisfied for γ 6= 1, we report in Table 6.7 the variation of condition number with the
weighting factor. Note that the optimal value γ ≈ 0.4 of the weighting parameter is
almost independent of the smoother (except for the condition number in case of point-
based Gauss-Seidel), and leads to a substantial decrease in the convergence rate (by
a factor of two or more for both Gauss-Seidel variants) and slightly less substantial
decrease in the condition number.
6.4.2 Multigrid implementation
We now consider the multigrid implementation of the RS algorithm. The convergence
behaviour of the method is investigated for V- and W-cycles [61], as well as for the
Krylov-based cycling strategy [49]. This latter is implemented as in Algorithm 3.2
from [48], with flexible conjugated gradient (FCG) acceleration at every level and with
t = 0 (that is, exactly two FCG iterations are performed). Since the choice of FCG
is relevant if the preconditioner is symmetric, we set ν1 = ν2 in what follows. The
resulting multigrid method is itself used on the finest grid as a preconditioner for the
FCG(1) method from [45] (which amounts to standard conjugate gradient method in
the case of V- and W-cycles).
In what follows we consider (xyz) prolongation with a direction-based Gauss-Seidel
smoother as the most interesting combination. This case is supplemented with the
Jacobi hybrid smoother to illustrate the effect of a less efficient smoothing scheme. The
iteration counts for the three cycling strategies are given in Table 6.8 for the periodic
case and in Table 6.9 for the Dirichlet one. In all cases the iterations counts are obtained
using 10 randomly chosen right hand sides (the same for three cycling strategies) and
136 Fourier Analysis of aggregation-based two-grid method for edge element
Jacobi direction-based GSnbr. ν1, ν2 = 1/2 ν1, ν2 = 1 ν1, ν2 = 1/2 ν1, ν2 = 1grids V W K V W K V W K V W K2 22 22 22 19 19 19 18-19 18-19 18-19 16 16 163 35-36 30 24-25 29 24 20-21 25-26 22 19 21 17-18 164 43-44 36-37 25 33 27-28 20-21 29 24 19 25 20 165 51 41 25-26 38 31 21-22 35 28-29 19 28-29 24 16
Table 6.8: Iteration counts for various cycling strategies; periodic boundary conditionsare considered; the finest grid corresponds to N = 33 (98304 edge unknowns).
Jacobi direction-based GSnbr. ν1, ν2 = 1/2 ν1, ν2 = 1 ν1, ν2 = 1/2 ν1, ν2 = 1grids V W K V W K V W K V W K2 23 23 23 20-21 20-21 20-21 19 19 19 16 16 163 39 32-33 28 32-33 27 23 28 23-24 20 22-23 19 17-184 51-52 42 29 40-41 33 23 31 26-27 20 23-24 19-20 17-185 57-58 45-47 30 42-43 34-35 23 31-32 27 20 24 20-21 17-18
Table 6.9: Iteration counts for various cycling strategies; Dirichlet boundary condi-tions are considered; the finest grid corresponds to N = 34 (101376 edge unknowns).
reducing the residual by a factor of 1010. Regarding the results for the Gauss-Seidel case
we conclude that, at least for the considered problem, the K-cycle multigrid converges
in almost the same number of iterations as the two-grid cycle implemented on the finest
grid. A slight increase is observed in the case of the Jacobi smoother, which is however
less pronounced than the one for V- and W-cycles.
6.5 Conclusion
We have performed the Fourier analysis of Reitzinger and Schoberl multigrid approach
on 3D curl-curl problems discretized with edge finite elements. We have shown that the
approach has level-independent convergence properties for various smoother configura-
tions and aggregates’ shapes. We have compared the results of the analysis with the
convergence rate of similar model problems and observed that the former give accurate
estimates of the later. We have observed that a few iterations of the Gauss-Seidel hybrid
smoother combined with the cubwise aggregation coarsening leads to a good compromise
between resource requirements and convergence speed. In multi-level setting, we have
observed that an almost level-independent convergence can be recovered when using
K-cycle.
List of Figures
1.1 An example of correction e smoothed by Gauss-Seidel scheme; (a) ini-tial correction (b) correction after 1 iteration (c) correction after 2 itera-tions. The corresponding linear system A was obtained by discretizationof constant-coefficient isotropic Poisson PDE (1.3) with Dirichlet bound-ary conditions on rectangular grid 33× 33. . . . . . . . . . . . . . . . . . 5
4.1 The dependence of ρ(E(J)MG) () and ρ(ETG) (×), as well as leftmost lower
bound (4.33) (∗) and rightmost upper bound (4.33) (+) , on the numberν of smoothing steps. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.1 Examples of (a) boxwise, (b) linewise and (c) L-shaped aggregation pat-terns. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.2 (a) general box aggregate situation with respect to discontinuities and (b)discontinuity nodes aggregated with white point nodes. . . . . . . . . . . . 107
5.3 Two potential aggregation strategies for the numerical example. . . . . . . 108
6.1 Edge neighbourhood. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
137
List of Tables
2.1 Convergence factor of V–cycle (for N0 = 2 and J = 6) and the correspond-ing bounds for ν = 1; (*) the quantity exists, but does not correspond tothe bound, since ω(1) > 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.2 Convergence factor of V–cycle (for N0 = 2 and J = 6) and the correspond-ing bounds for ν = 2; (*) the quantity exists, but does not correspond tothe bound, since ω(1) > 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.1 The estimates of main convergence parameters for ν = 1 and for differentdamping factors ωJac. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.2 The estimates of main convergence parameters for ν = 2 and for differentdamping factors ωJac. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.3 The values of main parameters for ε = 10−4 and for different problemsizes; the coarsest grid corresponds to N0 = 4. . . . . . . . . . . . . . . . . 52
3.4 The values of main parameters for different problem sizes; the coarsestgrid corresponds to N0 = 2. . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.1 Various prolongations coming from Table 6.2 in [68], with asymptoticalbehavior of tan2 φ(j) for small values of θ1 and θ2. . . . . . . . . . . . . . 78
4.2 The estimates of different terms involved in inequalities (4.33) for ν = 1, 2.Two-grid and V-cycle convergence factors are assessed considering J = 7and M0 = 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.1 The value of µD and of its upper bound (5.21) for different grid sizes. . . 975.2 The value of µD and of its upper bound (5.21) for different aggregation
strategies and for d = 10. . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.1 Convergence factor of a two-grid method with Jacobi hybrid smoother,estimated via Fourier analysis. . . . . . . . . . . . . . . . . . . . . . . . . 132
6.2 Convergence factor of a two-grid method with various Gauss-Seidel vari-ants of hybrid smoother, estimated via Fourier analysis. . . . . . . . . . . 132
6.3 Comparison of Fourier analysis and actual two-grid convergence factorsin the case of Jacobi hybrid smoother. . . . . . . . . . . . . . . . . . . . . 133
6.4 Comparison of Fourier analysis and actual two-grid convergence factorsfor point-based Gauss-Seidel hybrid smoother. . . . . . . . . . . . . . . . 133
6.5 Comparison of Fourier analysis and actual two-grid convergence factorsfor direction-based Gauss-Seidel hybrid smoother. . . . . . . . . . . . . . 134
139
140 List of Tables
6.6 Dependence of convergence rate on γ for (xyz) prolongation and varioussmoothers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.7 Dependence of condition number on γ for (xyz) prolongation and varioussmoothers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
6.8 Iteration counts for various cycling strategies; periodic boundary condi-tions are considered; the finest grid corresponds to N = 33 (98304 edgeunknowns). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
6.9 Iteration counts for various cycling strategies; Dirichlet boundary condi-tions are considered; the finest grid corresponds to N = 34 (101376 edgeunknowns). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
Bibliography
[1] Antonio Arico, Marco Donatelli, and Stefano Serra-Capizzano. V-cycle optimal
convergence for certain (multilevel) structured linear systems. SIAM J. Matrix
Anal. Appl., 26:186–214, 2004.
[2] Douglas N Arnold, Richard S Falk, and Ragnar Winther. Multigrid in H(div) and
H(curl). Numerische Mathematik, 85(2):197–217, 2000.
[3] O. Axelsson. Iterative Solution Methods. Cambridge University Press, Cambridge,
1994.
[4] R. Beck. Algebraic multigrid by components splitting for edge elements on simplicial
triangulations. Preprint SC 99-40, ZIB, December 1999.
[5] Rudolf Beck, Peter Deuflhard, Ralf Hiptmair, Ronald H.W. Hoppe, and Bar-
bara Wohlmuth. Adaptative multilevel methods for edge element discretizations
of Maxwell’s equations. Surveys on Mathematics for Industry, 8:271–312, 1999.
[6] A. Ben Israel and T. N. E. Greville. Generalized Inverses : theory and applications.
J. Wiley and Sons, New York, 1974.
[7] M. Benzi. Preconditioning techniques for large linear systems: A survey. J. Com-
putational Physics, 182:418–477, 2002.
[8] Tim Boonen. Multigrid algorithms for electromagnetic field computations. PhD
thesis, Katholieke Universiteit Leuven, Belgium, 2008.
[9] Tim Boonen, Jan Van lent, and Stefan Vandewalle. Local Fourier analysis of multi-
grid for the curl-curl equation. SIAM Journal on Scientific Computing, 30(4):1730–
1755, 2008.
[10] D. Braess. The convergence rate of a multigrid method with Gauss-Seidel relaxation
for the Poisson equation. In W. Hackbusch and U Trottenberg, editors, Multigrid
Methods, Lectures Notes in Mathematics No. 960, pages 368–386, Berlin Heidelberg
New York, 1982. Springer-Verlag.
141
142 Bibliography
[11] D. Braess. Towards algebraic multigrid for elliptic problems of second order. Com-
puting, 55:379–393, 1995.
[12] D. Braess. Finite Elements: Theory, Fast Solvers, and Applications in Solid Me-
chanics. Cambridge University Press, Cambridge, 1997.
[13] D. Braess and W. Hackbusch. A new convergence proof for the multigrid method
including the V -cycle. SIAM J. Numer. Anal., 20:967–975, 1983.
[14] James H. Bramble, Joseph E. Pasciak, Junping Wang, and Jinchao Xu. Convergence
estimates for multigrid algorithms without regularity assumptions. Math. Comp.,
57:23–45, 1991.
[15] A. Brandt, S. F. McCormick, and J. W. Ruge. Algebraic multigrid (amg) for sparse
matrix equations. In D. J. Evans, editor, Sparsity and its Application, pages 257–
284. Cambridge University Press, Cambridge, 1984.
[16] Achi Brandt. Algebraic multigrid theory: the symmetric case. Appl. Math. Comput.,
19:23–56, 1986.
[17] Achi Brandt. Rigorous quantitative analysis of multigrid, i: constant coefficients
two-level cycle with l2-norm. SIAM J. Numer. Anal., 31:1695–1730, 1994.
[18] M. Brezina, A. J. Cleary, R. D. Falgout, V. E. Henson, J. E. Jones, T. A. Man-
teuffel, S. F. McCormick, and J. W. Ruge. Algebraic multigrid based on element
interpolation (AMGe). SIAM J. Sci. Comput., 22:1570–1592, 2000.
[19] William L. Briggs, Van Emden Henson, and Stephen F. McCormick. A multigrid
tutorial. SIAM, Philadelphia, PA, 2000.
[20] V. E. Bulgakov. Multi-level iterative technique and aggregation concept with semi-
analytical preconditioning for solving boundary-value problems. Comm. Numer.
Methods Engrng., 9:649–657, 1993.
[21] T. Chartier, R. D. Falgout, V. E. Henson, J. Jones, T. Manteuffel, S. McCormick,
J. Ruge, and P. S. Vassilevski. Spectral AMGe (ρAMGe). SIAM J. Sci. Comput.,
25:1–26, 2004.
[22] Robert D. Falgout and Panayot S. Vassilevski. On generalizing the algebraic multi-
grid framework. SIAM J. Numer. Anal., 42:1669–1693, 2005.
[23] Robert D. Falgout, Panayot S. Vassilevski, and Ludmil T. Zikatanov. On two-grid
convergence estimates. Numer. Lin. Alg. Appl., 12:471–494, 2005.
[24] G. H. Golub and C. F. van Loan. Matrix Computations. The John Hopkins Uni-
versity Press, Baltimore, Maryland, 1996. Third ed.
Bibliography 143
[25] M. Griebel and P. Oswald. On the abstract theory of additive and multiplicative
Schwarz algorithms. Numer. Math., 70:163–180, 1995.
[26] W. Hackbusch. Multi-grid convergence theory. In W. Hackbusch and U Trottenberg,
editors, Multigrid Methods, Lectures Notes in Mathematics No. 960, pages 177–219,
Berlin Heidelberg New York, 1982. Springer-Verlag.
[27] W. Hackbusch. Multi-grid Methods and Applications. Springer, Berlin, 1985.
[28] M. R. Hestenes and E. Stiefel. Methods of conjugate gradients for solving linear
systems. J. Res. Nat. Bur. Standards, 49:409–436, 1952.
[29] R. Hiptmair. Multigrid method for Maxwell’s equations. SIAM Journal on Numer-
ical Analysis, 36(1):204–225 (electronic), 1999.
[30] J. Jones and B. Lee. A multigrid method for variable coefficient Maxwell’s equations.
SIAM Journal on Scientific Computing, 27(5):1689–1708, 2006.
[31] J. E. Jones and P. S. Vassilevski. AMGe based on element agglomeration. SIAM
J. Sci. Comput., 23:109–133, 2001.
[32] H. Kim, J. Xu, and L. Zikatanov. A multigrid method based on graph matching
for convection-diffusion equations. Numer. Lin. Alg. Appl., 10:181–195, 2003.
[33] Tzanio V. Kolev and Panayot S. Vassilevski. Amg by element agglomeration and
constrained energy minimisation interpolation. Numer. Lin. Alg. Appl., 13:771–788,
2006.
[34] Michael Lauzon and Sergei Treil. Common complements of two subspaces of a
Hilbert space. Journal of Functional Analysis, 212:500–512, 2003.
[35] B. Lee and C. Tong. A novel algebraic multigrid-based approach for Maxwell’s
equations. Preprint UCRL-JC-218750, 2006.
[36] R. B. Lehoucq, D. C. Sorensen, and C. Yang, editors. ARPACK Users’ Guide: Solu-
tion of Large-Scale Eigenvalue Problems with Implicitly Restarted Arnoldi Methods.
SIAM, Philadelphia, PA, 1998.
[37] Jan Mandel, Steve F. McCormick, and John W. Ruge. An algebraic theory for
multigrid methods for variational problems. SIAM J. Numer. Anal., 25:91–110,
1988.
[38] S. F. McCormick. Multigrid methods for variational problems: general theory for
the V -cycle. SIAM J. Numer. Anal., 22:634–643, 1985.
144 Bibliography
[39] Carl D. Meyer. Matrix Analysis and Applied Linear Algebra. SIAM, Philadelphia,
2000.
[40] K. W. Morton and D. F. Mayers. Numerical solution of partial differential equations:
an introduction. Cambridge University Press, Cambridge, 2005.
[41] A. C. Muresan and Y. Notay. Analysis of aggregation-based multigrid. SIAM J.
Sci. Comput., 30:1082–1103, 2008.
[42] S. Nakamura. Computational Methods in Engineering and Science. John Wiley &
Sons, New York, 1977.
[43] Michael K. Ng. Iterative methods for Toeplitz systems. Oxford University Press,
Oxford, 2004.
[44] Y. Notay. Algebraic analysis of two-grid methods: the nonsymmetric case. Numer.
Lin. Alg. Appl. published online in Wiley InterScience. DOI: 10.1002/nla.649, 2009.
[45] Y. Notay. Flexible conjugate gradients. SIAM J. Sci. Comput., 22:1444–1460, 2000.
[46] Y. Notay. Algebraic multigrid and algebraic multilevel methods: a theoretical
comparison. Numer. Lin. Alg. Appl., 12:419–451, 2005.
[47] Y. Notay. Convergence analysis of perturbed two-grid and multigrid methods. SIAM
J. Numer. Anal., 45:1035–1044, 2007.
[48] Y. Notay. An aggregation-based algebraic multigrid method. Electronic Trans.
Numer. Anal., 2009. To appear.
[49] Y. Notay and P. S. Vassilevski. Recursive Krylov-based multigrid cycles. Numer.
Lin. Alg. Appl., 15:473–487, 2008.
[50] P. Oswald. Multilevel Finite Element Approximation: Theory and Applications.
Teubner Skripte zur Numerik. Teubner, Stuttgart, 1994.
[51] P. Oswald. Subspace Correction Methods and Multigrid Theory, pages 533–572. In
Trottenberg et al. [61], 2001. Appendix A.
[52] S. Reitzinger and J. Schoberl. An algebraic multigrid method for finite element
discretizations with edge elements. Numerical Linear Algebra with Applications,
9(3):223–238, 2002.
[53] J. W. Ruge and K. Stuben. Algebraic multigrid (AMG). In S. F. McCormick,
editor, Multigrid Methods, volume 3 of Frontiers in Applied Mathematics, pages
73–130. SIAM, Philadelphia, PA, 1987.
Bibliography 145
[54] Y. Saad. Iterative Methods for Sparse Linear Systems. SIAM, Philadelphia, PA,
2003. Second ed.
[55] Y. Saad and H. van der Vorst. Iterative solution of linear systems in the 20-th
century. J. Comput. Appl. Math., 123:1–33, 2000.
[56] Gordon D. Smith. Numerical solution of partial differential equations: finite differ-
ence methods. Oxford University Press, Oxford, 1985.
[57] R. Stevenson. On the validity of local mode analysis of multi-grid methods. PhD
thesis, Utrecht University, 1990.
[58] K. Stuben. Algebraic multigrid (AMG): experiences and comparisons. Appl. Math.
Comput., 13:419–452, 1983.
[59] K. Stuben. An Introduction to Algebraic Multigrid, pages 413–532. In Trottenberg
et al. [61], 2001. Appendix A.
[60] K. Stuben and K. U. Trottenberg. Multigrid methods: Fundamental algorithms,
model problem analysis and applications. In W. Hackbusch and U. Trottenberg,
editors, Multigrid Methods, Lectures Notes in Mathematics No. 960, pages 1–176,
Berlin Heidelberg New York, 1982. Springer-Verlag.
[61] U. Trottenberg, C. W. Oosterlee, and A. Schuller. Multigrid. Academic Press,
London, 2001.
[62] Henk A. van der Vorst. Iterative Krylov methods for large linear systems. Cambridge
University Press, Cambridge, 2003.
[63] P. Vanek, J. Mandel, and M. Brezina. Algebraic multigrid based on smoothed
aggregation for second and fourth order problems. Computing, 56:179–196, 1996.
[64] Petr Vanek, Marian Brezina, and Radek Tezaur. Two-grid method for linear elas-
ticity on unstructured meshes. SIAM J. Sci. Comput., 21:900–923, 1999.
[65] P. S. Vassilevski. Multilevel Block Factorization Preconditioners. Springer, New
York, 2008.
[66] John L. Volakis, Arindam Chatterjee, and Leo C. Kempel. Finite Element Method
for Electromagnetics. IEEE Press, New York, 1998.
[67] P. Wesseling. An Introduction to Multigrid Methods. J. Wiley and Sons, Chichester,
1992.
[68] R. Wienands and W. Joppich. Practical Fourier Analysis for multigrid methods.
Chapman & Hall/CRC Press, Boca Raton, Florida, 2005.
146 Bibliography
[69] Roman Wienands and Cornelis W. Oosterlee. On three-grid Fourier analysis for
multigrid. SIAM J. Sci. Comput., 23:651–671, 2001.
[70] G. Wittum. Linear iterations as smoothers in multigrid methods: Theory with
applications to incomplete decompositions. Impact of Computing in Science and
Engineering, 1:180–215, 1989.
[71] G. Wittum. On the robustness of ILU-smoothing. SIAM J. Sci. Statist. Comput.,
10:699–717, 1989.
[72] J. Xu. Theory of Multilevel Methods. PhD thesis, Cornell University, 1989.
[73] J. Xu. Iterative methods by space decomposition and subspace correction. SIAM
Review, 34:581–613, 1992.
[74] J. Xu and L. T. Zikatanov. The method of alternating projections and the method
of subspace corrections in hilbert space. J. Amer. Math. Soc., 15:573–597, 2002.
[75] H. Yserentant. Old and new convergence proofs for multigrid methods. Acta Nu-
merica, 2:285–326, 1993.
[76] O.C. Zienkiewicz. Finite Elements: Introductory lectures on the finite element
method. McGraw-Hill, London, 1977.
[77] Ludmil T. Zikatanov. Two-sided bounds on the convergence rate of two-level meth-
ods. Numer. Lin. Alg. Appl., 15(5):439–454, 2007.